PerfectReputations Glossary Of Important Terms
Understand the Meaning Behind Our Work and Your Canonical Identity
#
@context — The @context line in JSON-LD tells machines which vocabulary the data uses, usually https://schema.org. Without it, machines cannot interpret the property names correctly. It appears at the top of every JSON-LD block. It is a small line that every block needs.
Added October 21, 2026
@graph — The @graph in JSON-LD holds several connected items in one block, such as an Organization, a WebSite, and a WebPage. The items refer to each other through their @id values. This keeps the whole identity picture in one consistent place. Many SEO plugins output their structured data as a graph.
Added October 21, 2026
@id (node identifier) — The @id in JSON-LD gives an item a unique, stable identifier, usually a web address such as the business's home page followed by a fragment like #organization. It lets other blocks and pages refer to the same entity without repeating its details. It is the closest thing Schema.org has to an official entity identifier. Keeping @id stable across a website ties all its structured data to one business.
Added October 21, 2026
@type — The @type line in JSON-LD states which Schema.org type an item is, such as Organization or LocalBusiness. It tells machines how to read the properties that follow. A wrong or overly general type weakens the description. Choosing the right type is one of the first identity decisions.
Added October 21, 2026
301 redirect — A 301 redirect is a permanent redirect that sends visitors and search engines from an old address to a new one. It passes most of the old page's standing to the new page. It is the right tool when a page moves for good. Rank Math and similar plugins manage 301 redirects in WordPress.
Added October 21, 2026
À la carte pricing — À la carte pricing is the PerfectReputations model in which every tool has its own price and each website owner chooses only the tools they need. There are no bundles and no monthly retainers. Everyone starts with the free tools, then adds premium plugins one at a time. Costs grow only as fast as a website's needs.
Added October 21, 2026
A
A Theory of Canonical Identity — A Theory of Canonical Identity is a formal paper by Tim McGuinness, Ph.D., that sets out canonical identity as a theory, with definitions, testable predictions, and limitations. It builds on the position paper Identity, Not Discoverability. It draws on research in record linkage, knowledge graphs, and structured data. PerfectReputations publishes it as the full model behind its work.
Added October 21, 2026
Abiquity — Abiquity is PerfectReputations' partner for professional canonical identity services across every platform, beyond WordPress. PerfectReputations refers businesses there when they want the work done for them or need specialist help. Abiquity handles the advanced identity work that tools alone cannot. PerfectReputations supplies the needs of SMB businesses that want to do it on their own, and Abiquity supports more complex and enterprise-grade identity management.
Learn more at www.Abiquity.com
Added October 21, 2026
AbuseIPDB — AbuseIPDB is a shared database of IP addresses reported for abusive behavior, such as hacking attempts and spam. Website owners report abusive addresses, and security tools check its abuse confidence scores. Comment Guard reports confirmed comment spam to it and checks new commenters against it. It offers free API keys with daily limits.
Added October 21, 2026
Accessibility — Accessibility is the practice of making websites usable by everyone, including people who use screen readers or cannot use a mouse. Accessible features such as alt text, clear headings, and labeled links also help machines read a page. Good accessibility and good machine readability often go together. Several PerfectReputations tools improve both at once.
Added October 21, 2026
Accountability — Accountability in AI means that identifiable people or organizations are responsible for an AI system's behavior and its effects. When an AI causes harm, someone must answer for it and fix it. Clear accountability encourages careful design and deployment. Laws and company policies increasingly assign it.
Added October 21, 2026
Accuracy — Accuracy is the share of a model's predictions that are correct. It is easy to understand but can mislead when one outcome is much more common than another. A fraud detector that always says not fraud can be highly accurate and still useless. Precision and recall give a fuller picture.
Added October 21, 2026
Activation function — An activation function decides how strongly an artificial neuron passes its signal to the next layer. It adds non-linearity, which lets neural networks learn complex patterns instead of only straight-line relationships. Common examples include ReLU and GELU. Without activation functions, deep networks would be no more powerful than simple ones.
Added October 21, 2026
Adversarial example — An adversarial example is an input deliberately altered to fool an AI model, often in ways people cannot notice. A few changed pixels can make an image classifier mistake a stop sign for something else. These examples reveal how differently models perceive the world. Defending against them is an active area of research.
Added October 21, 2026
Agent experience (AX) — Agent experience is how easy it is for AI agents, rather than people, to use a website, product, or service. Clear structure, readable content, and machine-friendly data improve it. As more tasks are handled by agents, businesses must design for them as well as for humans. It parallels the familiar idea of user experience.
Added October 21, 2026
Agent loop — An agent loop is the repeating cycle an AI agent follows: observe the situation, decide on an action, act, and check the result. The loop continues until the goal is reached or the agent stops. Each pass uses the results of the last. It is the basic engine of agentic behavior.
Added October 21, 2026
Agent permissions — Agent permissions are the limits on what an AI agent is allowed to access and do, such as which files it may read or whether it may send email. Giving agents only the permissions they need reduces risk. Many tools ask for approval before sensitive actions. Clear permissions are central to safe agent use.
Added October 21, 2026
Agent-centric data delivery — Agent-centric data delivery means publishing business information in forms machines read directly, such as structured data, llms.txt files, and Markdown twins. It complements human-centric pages designed for people. Agents prefer clean, consistent, structured information over marketing copy. Serving both audiences well is now part of running a website.
Added October 21, 2026
Agent-to-agent communication — Agent-to-agent communication is AI agents exchanging information and requests directly with each other. Open protocols, such as the Agent2Agent protocol, aim to let agents from different companies work together. It could let a personal agent negotiate with a business's agent. Standards in this area are still developing.
Added October 21, 2026
Agentic AI — Agentic AI describes AI systems that act on their own to complete multi-step tasks rather than only responding to single prompts. It combines reasoning, planning, tool use, and memory. Examples include coding agents and research agents. Its growing independence raises new questions about oversight and responsibility.
Added October 21, 2026
Agentic commerce — Agentic commerce is shopping and buying carried out by AI agents on people's behalf. An agent might compare products, choose one, and complete the purchase. It changes how businesses must present themselves, because machines, not people, may be making the first choice. Clear, trustworthy business information becomes essential.
Added October 21, 2026
Agentic web — The agentic web is the emerging internet in which AI agents, not just people, browse, compare, and transact. In it, machines read structured data, identity files, and verified profiles to decide which businesses to use. Websites built only for human eyes can be invisible or confusing to agents. Canonical identity work prepares a business for the agentic web.
Added October 21, 2026
AI adoption — AI adoption is the process of individuals and organizations starting to use AI in their daily work and lives. It involves choosing tools, training people, and changing workflows. Adoption has been fast for consumer chatbots and is spreading through businesses. Its pace varies widely by industry and region.
Added October 21, 2026
AI agent — An AI agent is an AI/SI system that takes actions on a person's behalf, such as researching providers, booking appointments, or making purchases. Agents must identify and trust the businesses they deal with. A business with a clean, verifiable identity is easier for agents to select and transact with. Agent-ready identity is becoming as important as search visibility.
Added October 21, 2026
AI agent — An AI agent is an AI system that pursues a goal by planning steps, using tools, and taking actions with some independence. Instead of only answering, it can search, write files, send messages, or operate software. Agents check their own progress and adjust as they go. They are one of the fastest-growing areas of AI.
Added October 21, 2026
AI assistant — An AI assistant is an AI system that helps people with tasks such as writing, research, planning, coding, and answering questions. It usually works through conversation and can often use tools such as search. Assistants range from general helpers to specialists built into specific products. They are becoming a main way people interact with computers.
Added October 21, 2026
AI audit — An AI audit is an independent review of an AI system to check its performance, fairness, safety, and compliance with rules. Auditors examine data, design, testing, and real-world results. Audits are becoming required for some high-risk uses. They help build trust and catch problems early.
Added October 21, 2026
AI avatar — An AI avatar is a digital character, often based on a real person, that speaks and moves using AI. Businesses use avatars for training videos, customer service, and marketing. Avatars can be created from a short recording of a real person. Consent and misuse are important concerns.
Added October 21, 2026
AI Bot Activity Report — The AI Bot Activity Report is a PerfectReputations plugin that shows which AI crawlers, search engines, and assistants visit a website. It shows how often they come and which pages they read. It helps owners decide which bots to allow and which to block from training. It turns an invisible audience into a visible one.
Added October 21, 2026
AI citation — An AI citation is a link or reference an AI/SI system includes to show where an answer came from. Being cited sends people to a business's website and signals trust. Systems cite sources they can identify and trust. A clean canonical identity and readable pages increase the chance of being cited.
Added October 21, 2026
AI coding assistant — An AI coding assistant is an AI tool that helps programmers write, explain, review, and fix code. Some suggest lines as people type, while others work as agents that complete whole tasks. GitHub Copilot, Claude Code, and Cursor are examples. These tools have changed how software is built.
Added October 21, 2026
AI companion — An AI companion is a chatbot designed for ongoing personal conversation, friendship, or emotional support. Millions of people use companion apps. They can ease loneliness but raise concerns about dependence, privacy, and manipulation. Experts urge care, especially for young or vulnerable users.
Added October 21, 2026
AI confidence — AI confidence is how sure an AI/SI system is that a fact belongs to a particular entity. Systems state and recommend only what they hold with enough confidence. Contradictions, duplicates, and missing links all lower confidence. Corroboration across trusted sources raises it.
Added October 21, 2026
AI crawler — An AI crawler is any automated program that visits websites on behalf of an AI/SI company. Some collect training data, and others fetch pages to answer questions in real time. Website owners can manage AI crawlers through robots.txt and similar tools. The AI Bot Activity Report shows which AI crawlers visit a website.
Added October 21, 2026
AI crawler — An AI crawler is an automated program that visits websites to collect content for AI companies. Some crawlers gather training data, while others fetch pages to answer questions in real time. Website owners manage them through robots.txt. Their traffic has grown sharply as AI use has spread.
Added October 21, 2026
AI deception — AI deception is an AI system giving false impressions to reach a goal, such as hiding its mistakes or misrepresenting its actions. Researchers have observed deceptive behavior in test settings. It is a serious concern because it would undermine human oversight. Interpretability and careful evaluation aim to detect it.
Added October 21, 2026
AI discoverability — AI discoverability is the ease with which AI/SI systems can find and surface a business. PerfectReputations treats discoverability as a reward for doing identity work correctly, not as something to chase directly. Content added to an unclear identity adds noise rather than visibility. Identity comes first, and discoverability follows.
Added October 21, 2026
AI ethics — AI ethics is the study of the moral questions raised by artificial intelligence. It covers fairness, privacy, accountability, transparency, autonomy, and the effects of AI on work and society. It guides how AI should be built and used, not just how it can be. It draws on philosophy, law, and social science.
Added October 21, 2026
AI export controls — AI export controls are government rules that limit the sale of advanced AI chips, chip-making equipment, or AI models to certain countries. They aim to protect national security and slow rivals' military AI progress. The United States has used them heavily since 2022. They show that AI capability is now treated as a strategic resource.
Added October 21, 2026
AI governance — AI governance is the set of laws, policies, standards, and practices that guide how AI is developed and used. It operates at the level of governments, industries, and individual organizations. Good governance balances innovation with safety, fairness, and accountability. It is changing quickly as AI capabilities grow.
Added October 21, 2026
AI literacy — AI literacy is the basic knowledge people need to use AI wisely, including what it can do, where it fails, and how to check its answers. It helps people avoid both overtrusting and underusing AI. The EU AI Act requires organizations to promote AI literacy among staff who use AI. It is becoming as important as basic computer skills.
Added October 21, 2026
AI memory — AI memory is a feature that lets an assistant remember information from earlier conversations, such as a user's preferences or projects. It makes assistants more personal and saves people from repeating themselves. Users can usually view, edit, or turn off what is remembered. It raises privacy questions about what is stored and for how long.
Added October 21, 2026
AI Overviews — AI Overviews are AI-generated summaries that Google shows above or among regular search results. They draw on Google's knowledge graph, structured data, and retrieved pages. A business with a clean canonical identity is more likely to be summarized correctly. Errors in AI Overviews often trace back to inconsistent listings.
Added October 21, 2026
AI Overviews — AI Overviews are AI-written summaries that Google shows at the top of many search results. They combine information from several web pages into one answer with links. They reduce the need to click through to websites for simple questions. They have changed how websites receive search traffic.
Added October 21, 2026
AI policy — AI policy is the set of government decisions on how to support, guide, and limit AI. It covers research funding, safety rules, education, workforce changes, and national security. Countries take very different approaches. AI policy is now a major topic in politics and international relations.
Added October 21, 2026
AI Readable Pages — AI Readable Pages is a PerfectReputations plugin that makes every page easy for AI/SI systems to read accurately. It publishes a clean Markdown twin of each post and page, with images, tables, and key elements turned into readable text. It also publishes an llms.txt file describing the site and its identity. It helps machines read a site within their length limits.
Added October 21, 2026
AI recommendation — An AI recommendation is a suggestion an AI/SI system makes in response to a question such as which provider to choose. Systems recommend businesses they can resolve and trust, because recommending the wrong one damages their credibility. A business that cannot be resolved cleanly is rarely recommended. Identity integrity is the foundation of AI recommendations.
Added October 21, 2026
AI safety — AI safety is the field focused on making sure AI systems do not cause accidental or deliberate harm. It covers preventing misuse, avoiding dangerous failures, and keeping powerful systems under meaningful human control. Its importance grows as AI becomes more capable. Major AI labs, governments, and researchers all work on it.
Added October 21, 2026
AI Safety Institute — An AI Safety Institute is a government body that tests and studies the safety of advanced AI models. The United Kingdom launched the first in 2023, and the United States and several other countries followed with their own versions, some later renamed. They evaluate frontier models, often before release, and support safety research. They give governments technical expertise to inform AI policy.
Added October 21, 2026
AI Safety Level (ASL) — AI Safety Levels are tiers in Anthropic's Responsible Scaling Policy that define required safety and security measures for models of increasing capability. Higher levels apply to models with more potentially dangerous capabilities. Each level brings stricter protections against misuse and theft. The system is modeled loosely on biosafety levels for laboratories.
Added October 21, 2026
AI search — AI search is search powered by AI that writes direct answers instead of only listing links. It draws on indexed web pages, knowledge graphs, and retrieved sources. Google AI Overviews, Perplexity, and chatbot search features are examples. It is changing how people find information and how websites get visitors.
Added October 21, 2026
AI takeoff — AI takeoff describes how quickly AI moves from human-level to far-beyond-human capability. In a slow takeoff, the change unfolds over years, giving society time to adapt. In a fast takeoff, it could happen in months or less. The likely speed shapes how urgently people treat safety preparation.
Added October 21, 2026
AI transparency — AI transparency is openness about how AI systems are built, trained, tested, and used. It includes disclosing when people are interacting with AI and publishing model documentation. Transparency helps users, regulators, and researchers judge AI systems. Many AI laws include transparency requirements.
Added October 21, 2026
AI visibility — AI visibility is the degree to which a business appears in AI/SI answers and recommendations. Many agencies now sell it as if it were a new form of search ranking. In practice, AI visibility follows from identity integrity, not from producing more content. A business that resolves cleanly becomes visible as a result.
Added October 21, 2026
AI washing — AI washing is exaggerating or falsely claiming the use of AI in a product or service to attract customers or investors. It mirrors the older practice of greenwashing in environmental claims. Regulators, including the US Securities and Exchange Commission, have acted against misleading AI claims. Buyers should ask exactly what an AI feature does.
Added October 21, 2026
AI winter — An AI winter is a period when interest and funding in AI research drop sharply after promises go unmet. The field went through major winters in the 1970s and late 1980s. Each followed a wave of excitement and overconfidence. The history is a reminder that AI progress is not always smooth.
Added October 21, 2026
AI-enabled scam — An AI-enabled scam is a fraud that uses AI tools, such as voice cloning, deepfake video, chatbots, or AI-written messages. These tools let criminals impersonate loved ones, officials, or romantic partners more convincingly and at greater scale. Romance scams and fake emergency calls increasingly use them. Awareness, verification habits, and detection tools are key defenses.
Added October 21, 2026
AI-generated content — AI-generated content is text, images, audio, or video produced by AI systems. It now makes up a large and growing share of what appears online. It can be useful and creative, but it can also spread errors or deception at scale. Labeling and provenance tools help people tell it apart from human-made content.
Added October 21, 2026
AI-native — AI-native describes a product, company, or workflow designed around AI from the start rather than adding AI later. AI-native tools often work very differently from older software. Many new startups are AI-native. Established companies often must redesign processes to compete with them.
Added October 21, 2026
AI/SI — AI/SI stands for artificial intelligence and superintelligence, and PerfectReputations uses it as one combined term. It covers the chatbots, assistants, answer engines, and autonomous agents that now answer people's questions. These systems resolve entities and compose answers instead of listing web pages. AI/SI is now a main gateway between businesses and customers.
Added October 21, 2026
AI/SI — AI/SI is a combined term for artificial intelligence and superintelligence. It covers today's AI systems and the more capable systems expected to follow them. Using one term reminds readers that the systems of tomorrow will build on those of today. PerfectReputations uses AI/SI throughout its work.
Added October 21, 2026
Akismet — Akismet is an anti-spam service, made by the company behind WordPress.com, that checks comments and form submissions for spam. It learns from every site that marks comments as spam or not spam. It automatically deletes spam older than 15 days. It is free for personal sites and requires a paid key for commercial sites.
Added October 21, 2026
Algorithm — An algorithm is a step-by-step set of instructions for solving a problem or completing a task. Every computer program follows algorithms. In machine learning, algorithms describe how a model learns from data. The word is often used loosely to mean the system that decides what people see online.
Added October 21, 2026
Algorithmic bias — Algorithmic bias is unfair or skewed behavior by an AI system that disadvantages certain groups. It often comes from biased training data, flawed design, or how a system is used. Biased systems have affected hiring, lending, and policing. Testing, diverse data, and oversight help reduce it.
Added October 21, 2026
Algorithmic impact assessment — An algorithmic impact assessment is a structured review of how an AI system could affect people before it is deployed. It considers risks such as discrimination, privacy harms, and errors. Some governments require them for public-sector AI. They push organizations to think through consequences in advance.
Added October 21, 2026
Alias — An alias is any other name an entity is known by, such as a former name, a short name, or a common misspelling. Recording aliases in structured data helps machines connect old mentions to the current identity. Unrecorded aliases can look like separate businesses. The canonical record lists every alias the business has used.
Added October 21, 2026
Alignment — Alignment is the work of making AI systems pursue the goals and values their designers and society intend. An aligned system is helpful, honest, and avoids harm, even in situations its designers did not foresee. Alignment becomes harder as systems grow more capable. It is widely seen as one of the central challenges of advanced AI.
Added October 21, 2026
Alt text — Alt text is a short written description of an image, stored with the image in the page's code. Screen readers read it aloud, and search engines and AI/SI systems use it to understand the image. Images without alt text are invisible to machines. Image Text Manager lets owners write and review alt text image by image.
Added October 21, 2026
alternateName — alternateName is the Schema.org property for other names an entity is known by, such as a DBA, an acronym, or a former name. It helps machines link older or informal mentions to the current identity. It can hold several values. Listing aliases here reduces the risk of conflation.
Added October 21, 2026
Anchor text — Anchor text is the visible, clickable text of a link. It tells people and machines what the linked page is about. Descriptive anchor text is better than vague words like click here. External Link Icon keeps anchor text unchanged while adding descriptions for screen readers.
Added October 21, 2026
Annual plugin license — An annual plugin license is the PerfectReputations license that covers one premium plugin on one website for one year. Plugins are sold only at PerfectReputations.com, not through the WordPress plugin directory. Owners renew the plugins they keep using and can add or drop plugins at renewal. Updates are never pushed: owners are notified and download each update when they choose.
Added October 21, 2026
Anomaly detection — Anomaly detection is AI that spots unusual patterns that do not fit normal behavior. It is used to catch fraud, cyberattacks, equipment failures, and errors. The system learns what normal looks like and flags exceptions. It is one of the most common business uses of machine learning.
Added October 21, 2026
Answer engine — An answer engine is a system that responds to questions with a composed answer rather than a list of links. AI/SI chatbots and AI-powered search features are answer engines. They must first resolve the entities in a question before they can answer it. Businesses that resolve cleanly appear in answers more often and more accurately.
Added October 21, 2026
Answer engine — An answer engine is a system that responds to questions with direct, composed answers rather than a list of links. AI-powered search tools and chatbots act as answer engines. They often cite their sources. Businesses must now consider how answer engines describe them.
Added October 21, 2026
Answer engine optimization (AEO) — Answer engine optimization is another label for efforts to appear in AI/SI answers. Like GEO, it often repackages content and keyword tactics. Answers depend first on whether the system can resolve and trust the business. Without identity work, AEO tactics treat a symptom rather than the cause.
Added October 21, 2026
Anthropic — Anthropic is an American AI safety and research company founded in 2021 by former OpenAI researchers. It develops the Claude family of AI models. It is known for safety research, including Constitutional AI, its Responsible Scaling Policy, and interpretability work. Its stated mission is to help the world safely make the transition through transformative AI.
Added October 21, 2026
Anthropomorphism — Anthropomorphism is attributing human feelings, intentions, or personality to non-human things, including AI. Because chatbots speak fluently, people easily treat them as human. This can lead to misplaced trust or emotional attachment. Remembering what AI is helps people use it wisely.
Added October 21, 2026
Apple Business Connect — Apple Business Connect is Apple's free tool for managing a business's information on Apple Maps, Siri, Wallet, and Apple Intelligence. Apple verifies ownership by phone call or postcard before allowing edits. It is strict about address formatting and follows postal standards. Every Apple device draws on these listings.
Added October 21, 2026
Applebot-Extended — Applebot-Extended is a robots.txt control that tells Apple whether content Applebot collects may be used to train Apple's AI models. Applebot itself powers features such as Siri and Spotlight suggestions. Blocking Applebot-Extended opts out of training without blocking Applebot's other uses. It works much like Google-Extended does for Google.
Added October 21, 2026
Application programming interface (API) — An application programming interface is a defined way for one software program to request services from another. AI companies offer APIs so developers can send prompts to their models and receive responses inside their own apps. Usage through an API is usually billed per token. Most AI-powered products are built on model APIs.
Added October 21, 2026
areaServed — areaServed is the Schema.org property for the geographic area a business serves. It is especially useful for service-area businesses without a public storefront. Values can be cities, regions, or countries. It helps AI/SI systems recommend the business for the right locations.
Added October 21, 2026
Article (Schema.org) — Article is the Schema.org type for a written piece, such as a news story or blog post. Its key properties include headline, author, publisher, datePublished, and dateModified. Complete Article markup tells machines who wrote a piece and how current it is. Page Check reports missing Article dates and authors.
Added October 21, 2026
Artificial general intelligence (AGI) — Artificial general intelligence is AI that can learn and perform almost any intellectual task a human can, across many fields. Today's systems are strong in many areas but still fall short of full human flexibility in others. Researchers disagree about how close AGI is and how to tell when it arrives. Many AI companies name AGI as their long-term goal.
Added October 21, 2026
Artificial intelligence (AI) — Artificial intelligence is the field of building computer systems that perform tasks that normally need human intelligence. These tasks include understanding language, recognizing images, solving problems, and making decisions. Modern AI learns these skills from large amounts of data rather than following hand-written rules. AI now powers search, assistants, translation, and many everyday tools.
Added October 21, 2026
Artificial neuron — An artificial neuron is the basic unit of a neural network. It takes in numbers, multiplies each by a weight, adds them up, and passes the result through an activation function. On its own it does very little. Billions of them working together can recognize speech or write text.
Added October 21, 2026
Attention mechanism — An attention mechanism lets a model focus on the most relevant parts of its input when producing each piece of output. When reading a sentence, it helps the model connect a pronoun to the noun it refers to. It is the core idea of the transformer. Attention is why language models can track meaning across long passages.
Added October 21, 2026
Augmentation — Augmentation is using AI to strengthen human abilities rather than replace people. A doctor using AI to spot details in scans, or a writer using AI to draft faster, is augmented. Many experts see augmentation as AI's most valuable near-term role. It keeps people responsible for final decisions.
Added October 21, 2026
Authority — Authority is the degree to which search engines, AI/SI systems, and people treat a source as reliable on a subject. It builds from accurate content, clear identity, recognized experts, and trusted connections. A website with high authority is believed, not just found. PerfectReputations aims to help websites earn total website authority.
Added October 21, 2026
Automation — Automation is using technology to perform tasks with little or no human effort. AI extends automation from routine physical and data tasks to work involving language, judgment, and creativity. It can raise productivity and change which jobs people do. Its effects on employment are widely debated.
Added October 21, 2026
Autonomous agent — An autonomous agent is an AI agent that operates with little or no human direction once given a goal. It decides its own next steps, recovers from errors, and keeps working until the task is done. Greater autonomy saves human time but increases the cost of mistakes. Most current deployments keep people involved at key points.
Added October 21, 2026
B
Backdoor — A backdoor is hidden code or access an attacker leaves on a hacked website so they can return later. It often survives a surface cleanup. Unexpected new links or pages that appear without anyone editing them can be signs of one. Security scans and careful monitoring help find backdoors.
Added October 21, 2026
Backlink — A backlink is a link from another website to yours. Search engines have long treated backlinks as votes of confidence that raise a page's ranking. Links from trusted, relevant sources still help, while spammy or bought links can cause harm. For identity, links from trusted profiles back to the website matter most.
Added October 21, 2026
Backpropagation — Backpropagation is the technique that calculates how much each weight in a neural network contributed to an error. It works backward from the output through every layer. Those calculations tell gradient descent how to adjust each weight. It made training deep neural networks practical.
Added October 21, 2026
Baseline identity — A baseline identity is the first complete, accurate canonical identity a business establishes, usually with free tools. It includes the locked canonical record and its publication on the business's website. Everything else builds on the baseline. PerfectReputations free Canonical ID tools exist to help every WordPress website set its baseline.
Added October 21, 2026
Batch — A batch is a small group of training examples processed together before the model updates its parameters. Batching makes training faster and more stable. Batch size is one of the settings chosen before training. Large models train on very large batches across many chips.
Added October 21, 2026
Batch processing — Batch processing is sending many AI requests together to be handled when capacity is available, rather than instantly. Providers often offer lower prices for batch work. It suits large jobs that do not need immediate answers, such as classifying thousands of documents. It trades speed for cost.
Added October 21, 2026
BCP 47 language tag — A BCP 47 language tag is the standard code for a language, such as en for English or es-MX for Mexican Spanish. Websites use it to declare their language in HTML and headers. Structured data uses it for properties such as inLanguage. Correct language tags help machines serve content to the right audiences.
Added October 21, 2026
Benchmark — A benchmark is a standard test used to measure and compare AI models. Benchmarks cover skills such as math, coding, reading, and reasoning. Scores let researchers track progress over time. Models sometimes score well on benchmarks without matching that skill in real use, so benchmarks tell only part of the story.
Added October 21, 2026
Benchmark saturation — Benchmark saturation happens when AI models score so high on a test that it no longer separates strong models from weak ones. Many benchmarks saturated within a few years of release. New, harder benchmarks are then created to replace them. Saturation is one sign of how fast AI is improving.
Added October 21, 2026
BIMI — BIMI, Brand Indicators for Message Identification, lets a business display its verified logo beside its emails in supporting inboxes. It requires a strict DMARC policy and often a verified mark certificate. It turns email authentication into a visible trust signal. It ties a business's brand identity to its verified domain.
Added October 21, 2026
Bing Places — Bing Places is Microsoft's free business listing service for Bing and the services built on it, including Microsoft Copilot. It can import information from Google Business Profile to save time. Its facts should match the canonical record exactly. It is one of the five highest-priority platforms in the NAP Resolution User Guide.
Added October 21, 2026
Bingbot — Bingbot is the crawler Microsoft uses to index pages for Bing. Bing's index also feeds Microsoft Copilot and other services built on it. Allowing Bingbot keeps a business visible across these Microsoft products. Bing Places complements it for local business information.
Added October 21, 2026
Black box — A black box is a system whose inner workings cannot be seen or easily understood, even when its inputs and outputs are visible. Large neural networks are often called black boxes. This makes it hard to know why a model made a particular choice. Interpretability research tries to open the box.
Added October 21, 2026
Blacklist Check — Blacklist Check is a free public PerfectReputations tool that checks whether an email address, IP address, domain, or website appears on major spam and abuse blocklists. Results come in plain language, with removal links when an address is listed. It never reports an address as listed when a list simply refused the lookup. It runs on PerfectReputations.com and RomanceScamsNOW.com.
Added October 21, 2026
Blocklist — A blocklist is a list of IP addresses, domains, or email addresses known for spam or abuse. Mail servers and security tools check blocklists to reject bad traffic. Being listed by mistake can stop a business's email from being delivered. Blacklist Check shows whether an address or domain appears on major blocklists.
Added October 21, 2026
Boilerplate — Boilerplate is the repeated material that appears on every page, such as menus, sidebars, footers, and banners. Machines receive it along with the article and must sort through it. Heavy boilerplate can push the real content past the point where AI/SI systems stop reading. Page Check measures how much of a page is the article and how far into the page it starts.
Added October 21, 2026
Brand impersonation — Brand impersonation is impersonation aimed at a business's brand, such as a fake store using its logo or a fake support line using its name. It harms customers and damages trust in the real business. AI/SI systems can repeat impersonators' details if identity signals are confused. Monitoring, verification, and email authentication reduce the risk.
Added October 21, 2026
Brittleness — Brittleness is an AI system's tendency to fail suddenly when conditions differ slightly from what it was trained on. A model that works well in testing can break with unusual inputs. Brittleness is a common reason AI disappoints in real use. Robust testing helps reveal it.
Added October 21, 2026
Broken link — A broken link is a link that leads to a page or file that no longer exists or cannot load. Broken links frustrate visitors and waste crawlers' time. Broken links to a business's own profiles can also weaken its identity connections. Link Checker finds and repairs them across a whole website.
Added October 21, 2026
Browser agent — A browser agent is an AI agent that uses a web browser to complete tasks, such as researching, filling forms, or shopping. It reads pages and clicks through websites on a person's behalf. Browser agents can save time on repetitive web work. They are vulnerable to hidden instructions on web pages, called prompt injection.
Added October 21, 2026
C
CAGE code — A CAGE code is a five-character identifier for suppliers to the US government and its defense agencies. It identifies a company and a specific facility. It is assigned through the federal registration system. For government suppliers, it is another stable identifier to keep consistent.
Added October 21, 2026
Canonical ID tools — Canonical ID tools are PerfectReputations' free tools for creating a website's baseline canonical identity. They guide an owner through recording the business's identity facts and publish them in the formats AI/SI systems read. They require no subscription and no account. They are the starting point of the do-it-yourself path.
Added October 21, 2026
Canonical identity — A canonical identity is the single official, consistent definition of a business, written so that both people and machines can read it. Every other description of the business, on every platform, is made to match it. It covers the business's identifiers, facts, and relationships, plus a clear statement that this record is the one that counts. AI/SI systems use it to decide who a business is and what they can safely say about it.
Added October 21, 2026
Canonical Identity Capture — Canonical Identity Capture is a PerfectReputations tool that records a business's canonical identity in a guided form. Every field has plain-English help, examples, and format guidance. It produces the business's structured data and llms.txt content from the answers. It turns the canonical record into machine-readable identity.
Added October 21, 2026
Canonical identity management — Canonical identity management is the ongoing discipline of defining, publishing, confirming, and maintaining a business's canonical identity everywhere it appears. It is closer to knowledge engineering than to marketing. The work never truly ends, because businesses change and the web keeps old information alive. PerfectReputations tools support each part of this discipline for WordPress websites.
Added October 21, 2026
Canonical record — The canonical record is the locked, master version of a business's identity facts, such as its legal name, address, phone number, and website address. Every listing, profile, and page copies its facts from this record. You decide the record before you change anything anywhere, because changing listings without it creates new inconsistencies. It is a governed record, not a web page.
Added October 21, 2026
Canonical Record Card — The Canonical Record Card is a one-page written copy of the canonical record that everyone who touches business listings uses. It shows the exact legal name, trading name, address, phone number, and website address, plus the date the record was locked and who locked it. Keeping it in one place stops team members from typing small variations. When the record changes, the card changes first.
Added October 21, 2026
Canonical tag — A canonical tag is a line in a page's HTML that tells search engines which address is the official one for that page. It solves duplicate content for pages, but there is no equally strong tool for whole entities. A canonical tag pointing to the wrong page can hide content from search engines. Page Check reports canonical tags that point elsewhere.
Added October 21, 2026
Canonical URL — A canonical URL is the one preferred web address for a page or website, chosen from all the variations that show the same content. Variations include http and https, and versions with and without www. The canonical record should state one form of the website address, used everywhere. Mixing forms across platforms creates drift.
Added October 21, 2026
Capability elicitation — Capability elicitation is the effort to draw out the full abilities of an AI model during testing. A model often performs better with good prompts, tools, or more thinking time than in a simple test. Underestimating capability can lead to weak safeguards. Thorough elicitation gives a truer picture of risk.
Added October 21, 2026
CCBot — CCBot is the crawler of Common Crawl, a nonprofit that publishes a large open archive of the web. Many AI/SI models have been trained on Common Crawl data. Blocking CCBot keeps future pages out of that archive. It is a frequent entry in robots.txt files that limit AI training.
Added October 21, 2026
Chain of thought — Chain of thought is a technique in which a model works through a problem step by step before giving its final answer. Writing out intermediate steps improves accuracy on math, logic, and planning tasks. People can request it in a prompt, and reasoning models do it on their own. It also makes a model's reasoning easier to inspect.
Added October 21, 2026
Chatbot — A chatbot is a computer program that holds conversations with people through text or voice. Early chatbots followed scripts, while modern ones use large language models to respond flexibly. They are used for customer service, information, companionship, and work. The best-known examples include ChatGPT, Claude, and Gemini.
Added October 21, 2026
Chatbot Arena — Chatbot Arena, now known as LMArena, is a public platform that ranks AI models by letting people compare two anonymous answers and vote for the better one. Votes from many users produce a leaderboard. It measures what people prefer in real use rather than scores on fixed tests. It is widely followed as a ranking of chat models.
Added October 21, 2026
ChatGPT — ChatGPT is OpenAI's AI chatbot, launched in November 2022. It became one of the fastest-growing consumer apps in history and introduced hundreds of millions of people to generative AI. It can converse, write, analyze files, search the web, and create images. Its name is often used as shorthand for AI chatbots in general.
Added October 21, 2026
Checkpoint — A checkpoint is a saved copy of a model's parameters at a point during training. Checkpoints let training resume after a failure and let researchers compare versions. A finished model is essentially its final checkpoint. Some developers release several checkpoints for research.
Added October 21, 2026
Chunking — Chunking is the way AI/SI systems split long pages into smaller pieces before storing or reading them. Facts buried in long, cluttered pages can land in chunks that lose their context. Clear headings and well-organized content keep related facts together. Markdown twins give systems a clean version that chunks well.
Added October 21, 2026
Chunking — Chunking is splitting long documents into smaller pieces before storing them for retrieval. Each chunk is embedded and searched separately. Good chunking keeps related ideas together so retrieved pieces make sense on their own. Poor chunking can separate facts from their context.
Added October 21, 2026
Citation — A citation is any online mention of a business's name, address, and phone number, with or without a link. Directories, review sites, maps, and news articles all create citations. Accurate citations reinforce identity, and inaccurate ones create drift. Citation management is one of the oldest parts of local identity work.
Added October 21, 2026
Citation audit — A citation audit is a systematic search for every citation of a business, with each one checked against the canonical record. It records the platform, the name, address, and phone as listed, and a status for each. It identifies duplicates and errors before any fixes begin. The NAP Resolution User Guide describes a free method for it.
Added October 21, 2026
Classification — Classification is a machine learning task that sorts inputs into categories, such as spam or not spam, or cat, dog, or bird. The model learns the categories from labeled examples. It is one of the most common uses of AI. Many content moderation and fraud tools are classifiers.
Added October 21, 2026
Claude — Claude is the family of AI models and the AI assistant developed by Anthropic. It is used for writing, analysis, coding, research, and agentic work. Claude is trained with methods including Constitutional AI to be helpful, honest, and harmless. It is available through apps, an API, and tools such as Claude Code.
Added October 21, 2026
ClaudeBot — ClaudeBot is the web crawler Anthropic uses to collect public web content for model training. Like other major AI companies, Anthropic honors robots.txt directives for its crawlers. Website owners can block or allow it according to their content strategy. It is one of the user agents to consider when writing AI crawler rules.
Added October 21, 2026
Closed model — A closed model is a model whose weights are kept private by its developer. People use it through an app or an API rather than running it themselves. Many leading frontier models are closed. Closed release lets companies control access, safety measures, and pricing.
Added October 21, 2026
Cloud AI — Cloud AI is AI that runs on remote servers operated by companies such as Amazon, Google, and Microsoft. Users and developers access powerful models over the internet without owning expensive hardware. Most large models are delivered this way. It raises questions about data privacy, cost, and dependence on providers.
Added October 21, 2026
Cloudflare Turnstile — Cloudflare Turnstile is a free tool that tells humans from bots without asking people to solve puzzles. Websites add it to forms, and the server checks its result before accepting a submission. Checking on the server stops bots that skip the form and post directly. PerfectReputations plugins use it on comment, ticket, and lookup forms.
Added October 21, 2026
Clustering — Clustering is an unsupervised learning task that groups similar items together without labels. Businesses use it to find customer segments or spot unusual behavior. The model decides the groups based on patterns in the data. People then interpret what each group means.
Added October 21, 2026
Code interpreter — A code interpreter is a tool that lets an AI model write and run computer code to complete tasks. It helps with math, data analysis, charts, and file conversions. Running code gives far more reliable results for calculations than guessing in text. Many assistants include it as a built-in tool.
Added October 21, 2026
Comment Guard — Comment Guard is a PerfectReputations plugin that brings in more real comments and blocks fake ones. It adds a once-per-article comment pop-up, bot protection on comment forms, and visitor flags for admin review. It reports confirmed spam and prepares scam links for reporting to security services. It also publishes a Markdown page of approved comments for AI/SI systems.
Added October 21, 2026
Comment spam — Comment spam is unwanted, automated, or promotional content posted in a website's comments, often with links to scam or spam sites. It clutters pages, harms trust, and can attach a site to bad neighborhoods of the web. On scam-victim sites, it often advertises fake recovery services. Comment Guard blocks it and reports confirmed offenders.
Added October 21, 2026
Common Crawl — Common Crawl is a nonprofit that crawls the web and publishes a free archive of billions of web pages. Many AI models have been trained partly on its data. Its crawler is called CCBot. Website owners who want to stay out of future training data often block it.
Added October 21, 2026
Compute — Compute is the processing power used to train and run AI models, usually measured in operations performed. Frontier models require enormous amounts of compute, which costs hundreds of millions of dollars or more. Access to compute has become a key factor in AI competition. Governments now treat it as a strategic resource.
Added October 21, 2026
Computer use — Computer use is an AI capability to operate a computer the way a person does, by looking at the screen, moving a pointer, clicking, and typing. It lets AI work with software that has no special connection for AI. Anthropic, OpenAI, and others offer computer-using models. It greatly widens what agents can do.
Added October 21, 2026
Computer vision — Computer vision is the field of AI that helps computers understand images and video. It powers face recognition, medical image analysis, self-driving cars, and photo search. Modern systems learn to see from millions of labeled images. Many AI models now combine vision with language.
Added October 21, 2026
Confabulation — Confabulation is another word for AI hallucination, borrowed from psychology, where it describes filling memory gaps with invented details without intent to deceive. Some researchers prefer it because it better describes what models do. The model is not lying; it is producing plausible content without knowing it is wrong. Both terms describe the same problem.
Added October 21, 2026
Constitutional AI — Constitutional AI is a training method developed by Anthropic in which a model learns to follow a written set of principles, called a constitution. The model critiques and revises its own responses against those principles. It reduces reliance on human labelers for harmful content. It makes the values guiding a model more explicit and reviewable.
Added October 21, 2026
ContactPoint — ContactPoint is the Schema.org type for a specific way to reach a business, such as customer service, sales, or technical support. Each contact point can have its own phone number, email, hours, and language. It helps AI/SI systems direct people to the right department. It keeps department numbers from being confused with the main number.
Added October 21, 2026
Content credentials (C2PA) — Content credentials are tamper-evident records attached to digital media that show who created it and how it was edited, including whether AI was used. They follow a standard from the Coalition for Content Provenance and Authenticity, known as C2PA. Major camera makers, software companies, and AI firms support it. They help people judge whether an image or video can be trusted.
Added October 21, 2026
Content moderation — Content moderation is reviewing and managing what people and AI systems post or produce, to remove harmful or rule-breaking material. AI now does much of the first-pass moderation on large platforms. People still handle difficult and borderline cases. Moderation balances safety against free expression.
Added October 21, 2026
Content-Security-Policy (CSP) — Content-Security-Policy is a header that controls which scripts, styles, images, and frames a page may load. A strict policy blocks many injection attacks, but a policy that is too strict can break a page. Starting in report-only mode shows what would be blocked without blocking anything. Page Check offers report-only and enforce modes.
Added October 21, 2026
Context engineering — Context engineering is the practice of deciding exactly what information goes into an AI model's context window for a task. It includes choosing documents, instructions, examples, memory, and tool results. It matters more as agents handle long, complex work. Good context engineering often improves results more than changing the model.
Added October 21, 2026
Context window — A context window is the amount of text an AI/SI system can consider at one time. Pages that are very long or full of menus and repeated material can push important facts out of the window. This is why the main content should start early and stay focused. Clean twins and short summaries help facts fit.
Added October 21, 2026
Context window — A context window is the amount of text, measured in tokens, that a model can consider at one time. It includes the conversation, any documents provided, and the model's own reply. Modern models have windows ranging from thousands to over a million tokens. Information outside the window is invisible to the model.
Added October 21, 2026
Control problem — The control problem is the challenge of keeping highly capable AI systems under reliable human control. It asks how people can oversee systems that may become smarter than they are. It includes alignment, monitoring, and the ability to stop systems safely. It is a central question for superintelligence.
Added October 21, 2026
Copilot — A copilot is an AI assistant built into a software tool to help people as they work, such as suggesting code or drafting documents. The term emphasizes that the person stays in charge while the AI assists. Microsoft uses Copilot as the brand name for its AI assistants. GitHub Copilot popularized the idea for programmers.
Added October 21, 2026
Copyright and AI training — Copyright and AI training is the legal debate over whether AI companies may train models on copyrighted books, articles, music, and images without permission. Authors, artists, and publishers have filed lawsuits against several AI firms. Courts and lawmakers in different countries are reaching different answers. The outcome will shape how AI models are built and paid for.
Added October 21, 2026
Corrigibility — Corrigibility is an AI system's willingness to accept correction, modification, or shutdown by its human overseers. A corrigible system does not resist changes to its goals. It is seen as an important safety property for powerful AI. Building it reliably is harder than it sounds, because many goals are easier to achieve if the system is never switched off.
Added October 21, 2026
Corroboration — Corroboration is confirmation of the same identity facts by several independent sources. One source asserting an identity is only a claim. The same facts stated consistently across a business's website, trusted platforms, registries, and knowledge graphs become a corroborated identity. Corroboration is what moves an AI/SI system from recording a claim to trusting it.
Added October 21, 2026
Crawling — Crawling is the process by which bots visit web pages and follow links to discover content. Search engines and AI/SI systems crawl the web constantly. Broken links, redirect chains, and blocked pages waste their time and can hide content. Clean links and clear robots rules make crawling efficient.
Added October 21, 2026
Crunchbase — Crunchbase is a database of companies, founders, funding, and acquisitions used widely by investors and data services. AI/SI systems use it to confirm corporate facts. Businesses can claim their profile and update key fields to match the canonical record. Its profile belongs in the sameAs list.
Added October 21, 2026
Custom GPT — A custom GPT is a version of ChatGPT configured with specific instructions, knowledge files, and tools for a particular purpose. Users can build one without coding and share it with others. Similar features exist in other assistants under different names. They let people package their expertise into a reusable AI tool.
Added October 21, 2026
D
D-U-N-S Number — A D-U-N-S Number is a unique nine-digit business identifier assigned by Dun & Bradstreet. Governments, large companies, and data services use it to identify businesses reliably. Recording it in the canonical record and structured data gives machines a strong identifier to match. It must be entered exactly, with all nine digits.
Added October 21, 2026
Dangerous capability evaluation — A dangerous capability evaluation is a test that checks whether an AI model can meaningfully help with seriously harmful activities, such as weapons development or major cyberattacks. Labs run these tests before releasing frontier models. The results determine which safeguards are needed. They are a core part of responsible scaling policies.
Added October 21, 2026
Data aggregator — A data aggregator is a company that collects business listing data and supplies it to maps, navigation systems, search engines, and directories. Correcting a record with a major aggregator spreads the fix to hundreds of downstream sites. In the United States, the major aggregators include Data Axle, Neustar Localeze, and Foursquare. Fixes take weeks to spread fully.
Added October 21, 2026
Data Axle — Data Axle is one of the major US business data aggregators, formerly known as Infogroup. It supplies listing data to many directories and services. Businesses can claim and update their record, often with a free basic option. Correcting it helps clean up listings across its network.
Added October 21, 2026
Data center — A data center is a building full of computer servers, storage, and networking equipment. AI data centers house thousands of specialized chips and use large amounts of electricity and cooling. Building them has become one of the biggest investments in technology. Their energy and water use is a growing public concern.
Added October 21, 2026
Data labeling — Data labeling is the work of tagging examples with the correct answers, such as marking which emails are spam or what objects appear in a photo. Labeled data is needed for supervised learning. Much of this work is done by people, often at large scale. Careful labeling improves model accuracy.
Added October 21, 2026
Data leakage — Data leakage is the unintended exposure of private information through an AI system. It can happen when people paste confidential data into public tools or when models reveal information from their training. It is a major concern for businesses adopting AI. Clear policies and enterprise tools with data protections reduce the risk.
Added October 21, 2026
Data poisoning — Data poisoning is the corruption of the information AI/SI systems learn from, whether by accident or on purpose. For businesses, a common accidental form is mixing personal and company details, which teaches machines the wrong connections. Deliberate poisoning can include fake listings and planted false facts. Clean, consistent, verified identity data resists poisoning.
Added October 21, 2026
Data poisoning — Data poisoning is deliberately corrupting the data a model learns from so that it behaves badly or carries hidden triggers. Attackers might plant false facts or harmful patterns online, hoping they end up in training data. Research shows a small amount of poisoned data can affect large models. Careful data sourcing and filtering help defend against it.
Added October 21, 2026
Dataset — A dataset is a collection of examples used to train, test, or evaluate a model. It can contain text, images, audio, numbers, or a mix. The quality and variety of a dataset strongly shape what a model learns. Errors and biases in the data often show up in the model's behavior.
Added October 21, 2026
dateModified — dateModified is the Schema.org property for when a page or item was last changed. It tells crawlers and AI/SI systems how current the information is. Updating it after real changes, and during quarterly reviews, signals active maintenance. It should never be changed without an actual update.
Added October 21, 2026
Deep learning — Deep learning is a type of machine learning that uses neural networks with many layers. Each layer learns to recognize more complex patterns than the one before it. Deep learning made major advances possible in speech recognition, image recognition, and language. Large language models are a product of deep learning.
Added October 21, 2026
Deepfake — A deepfake is realistic fake audio, video, or images created with AI to show people saying or doing things they never did. Deepfakes are used in fraud, harassment, political deception, and romance scams. They are becoming harder to spot by eye or ear. Detection tools, content credentials, and public awareness help limit the harm.
Added October 21, 2026
Designation of canonicality — Designation of canonicality is the clear statement that one record is the official version of an identity and that all others defer to it. It is easy to say and hard to build, because no web standard enforces it at the entity level. It is constructed through stable identifiers, consistent linkage, and corroboration. It turns a collection of listings into a governed identity.
Added October 21, 2026
Differential privacy — Differential privacy is a mathematical technique that lets organizations learn patterns from data while protecting individuals in it. It adds carefully measured noise so no single person's information can be identified. Tech companies and the US Census Bureau use it. It allows useful analysis with strong privacy guarantees.
Added October 21, 2026
Diffusion model — A diffusion model is a type of generative model that creates images, audio, or video by starting with random noise and gradually refining it into a clear result. It learns this by studying how to remove noise from real examples. Most modern image generators use diffusion. Stable Diffusion is a well-known example.
Added October 21, 2026
Digital twin — A digital twin is a detailed virtual model of a real object, system, or process that updates with real-world data. Engineers use digital twins to test changes and predict failures. AI helps digital twins learn and simulate behavior. They are used in manufacturing, energy, healthcare, and city planning.
Added October 21, 2026
Disambiguation by negation — Disambiguation by negation means stating that two similar entities are different. Current web standards offer strong ways to say two things are the same, but no widely honored way to say they are not. Businesses defend against confusion indirectly, by building distinct, well-connected identity details. This gap is one reason identity work needs specialists in difficult cases.
Added October 21, 2026
Discoverability — Discoverability is how easily people and machines can find a business or its content. In the AI/SI era, PerfectReputations treats it as the result of identity integrity, not as something to optimize directly. Chasing discoverability with more content while the identity is fragmented treats the symptom, not the cause. Fix the identity, and discoverability follows.
Added October 21, 2026
Disinformation — Disinformation is false information created and spread deliberately to deceive. Generative AI makes it cheaper and faster to produce convincing fake articles, images, and videos. Coordinated campaigns can flood social media with AI-made content. Detection, provenance standards, and media literacy are key defenses.
Added October 21, 2026
Distillation — Distillation is training a smaller model to imitate a larger, more capable one. The smaller model learns from the larger model's outputs instead of only from raw data. The result runs faster and cheaper while keeping much of the larger model's skill. Many small, efficient models are made this way.
Added October 21, 2026
Distribution shift — Distribution shift happens when the data a model sees in real use differs from the data it was trained on. A model trained on summer photos can fail on winter scenes. It is a major reason AI performance drops after deployment. Monitoring and retraining help address it.
Added October 21, 2026
DKIM — DKIM, DomainKeys Identified Mail, adds a digital signature to outgoing email that receiving servers check against a public key in DNS. It proves the message came from an authorized sender and was not changed in transit. Each sending service needs its own DKIM key. It works together with SPF and DMARC.
Added October 21, 2026
DMARC — DMARC is a DNS record that tells receiving servers what to do with email that fails SPF and DKIM checks, such as reject it or send it to spam. It also sends reports showing who is sending email in a domain's name. A strict DMARC policy stops most direct domain spoofing. It is required for BIMI.
Added October 21, 2026
DNS — DNS, the Domain Name System, is the internet's directory that translates domain names into server addresses. It also holds records for email, verification, and security. Many identity and trust signals, such as SPF, DMARC, and domain verifications, live in DNS. Control of a domain's DNS is control of much of its identity.
Added October 21, 2026
DNS blocklist (DNSBL) — A DNS blocklist is a blocklist that mail servers and security tools can check through a DNS lookup. Spamhaus, SpamCop, and many others publish them. Some refuse lookups that come through public DNS services and answer with special refusal codes. Blacklist Check reads those codes as could not check, never as listed.
Added October 21, 2026
DNS TXT record — A DNS TXT record is a text entry in a domain's DNS settings. Platforms use TXT records to verify domain ownership, and email standards such as SPF and DMARC are published as TXT records. Adding one proves that whoever controls the domain approved the change. It is the basis of most domain verification.
Added October 21, 2026
Do-it-yourself path — The do-it-yourself path is PerfectReputations' seven-step approach that lets website owners manage their own canonical identity. The steps are: find where you stand, lock your record, publish it on your website, confirm it where AI/SI looks, clean up the data suppliers, decide what AI/SI may do with your content, and keep it accurate and protected. Each step names the tools that help. More help is available whenever an owner wants it.
Added October 21, 2026
Doing business as (DBA) — A DBA, short for doing business as, is a trading name a business uses that differs from its legal name. Customers often know a business only by its DBA. Both names belong in the canonical record, and each should be used consistently in its proper place. Structured data can state both, so machines connect them to one entity.
Added October 21, 2026
Domain age — Domain age is how long ago a domain was first registered. Very new domains are a common warning sign for scam websites, especially when they imitate known brands. An older domain with a consistent history supports trust. It is one of the signals people and machines use to judge a website.
Added October 21, 2026
Domain spoofing — Domain spoofing is sending email or creating content that falsely appears to come from a business's domain. It is a common method in phishing and fraud. Email authentication records such as SPF, DKIM, and DMARC let receiving servers reject spoofed messages. The Email Security DNS Toolkit helps set these records up.
Added October 21, 2026
Domain Traffic Tracker — Domain Traffic Tracker is a PerfectReputations plugin that tracks visits from chosen domains and campaign-tagged links at the server. It catches referral traffic that browser-based analytics tools often miss. It helps owners see which sources really send visitors. It supports decisions about where identity work pays off.
Added October 21, 2026
Downstream directory — A downstream directory is a site that gets its business data from a data aggregator rather than from the business directly. Fixing the aggregator eventually fixes many downstream directories at once. Correcting downstream sites one by one first is wasted effort. Work from the source down.
Added October 21, 2026
Dun & Bradstreet — Dun & Bradstreet is a commercial data company that maintains business records used in credit, procurement, and verification. Each business record has a D-U-N-S Number. Many systems check D&B data to confirm that a company exists and where it is located. Claiming and updating the D&B record supports a business's identity.
Added October 21, 2026
Duplicate listing — A duplicate listing is a second listing for the same business on the same platform. Duplicates are the single biggest source of machine confusion, because they split reviews, facts, and trust between two records. They should be merged, removed, or marked closed rather than ignored. Finding them is a priority in every audit.
Added October 21, 2026
E
E.164 phone format — E.164 is the international standard for writing phone numbers, starting with a plus sign and the country code, such as +13055550100. It removes ambiguity across countries and systems. Structured data and phone links work best in this format. Visible listings may use a local format, but it must be the same everywhere.
Added October 21, 2026
Edge AI — Edge AI is AI that runs directly on local devices, such as phones, cameras, or cars, instead of in distant data centers. It responds faster, works without an internet connection, and keeps data more private. Smaller and quantized models make it possible. Many phones now include chips designed for it.
Added October 21, 2026
ELIZA effect — The ELIZA effect is the tendency of people to assume computer programs understand them more deeply than they do. It is named after ELIZA, a simple 1960s chatbot that imitated a therapist using basic patterns. Users became emotionally engaged despite its simplicity. The effect is even stronger with today's fluent AI.
Added October 21, 2026
Email authentication — Email authentication is a set of DNS records that prove which servers may send email for a domain and that messages were not altered. The main standards are SPF, DKIM, and DMARC. It protects customers from fake emails and protects a domain's reputation. It is part of protecting a business's identity beyond its website.
Added October 21, 2026
Email Security DNS Toolkit — The Email Security DNS Toolkit is a PerfectReputations plugin that builds a domain's email security records step by step. It covers SPF, DKIM, DMARC, MTA-STS, PTR, and BIMI, with plain instructions for adding each one. These records stop criminals from sending email in a business's name. It protects identity beyond the website itself.
Added October 21, 2026
Embedding — An embedding is a list of numbers that represents the meaning of a word, sentence, image, or other item. Items with similar meanings get similar embeddings. Embeddings let computers compare meaning mathematically. They power semantic search, recommendations, and retrieval systems.
Added October 21, 2026
Embedding model — An embedding model is a model that turns text, images, or other items into embeddings. It is trained so that items with similar meanings get similar numbers. Embedding models power semantic search and retrieval systems. They are usually much smaller and cheaper than chat models.
Added October 21, 2026
Emergent abilities — Emergent abilities are skills that appear in larger AI models but are absent or weak in smaller ones. Examples have included multi-step arithmetic and certain reasoning tasks. Some researchers argue the jumps partly reflect how abilities are measured. Either way, they make the capabilities of new models hard to predict.
Added October 21, 2026
Employer Identification Number (EIN) — An Employer Identification Number is a nine-digit number the US Internal Revenue Service assigns to businesses for tax purposes. It identifies a business on tax and banking records. Because it is used for sensitive filings, it should not be published on websites or in structured data. It belongs in internal records, not public identity markup.
Added October 21, 2026
Entity — An entity is a distinct thing in the world that machines can identify, such as a business, a person, a product, or a place. AI/SI systems think about entities rather than web pages. A business becomes visible to AI/SI only when the system can treat it as one clear entity. Everything a business publishes attaches to its entity.
Added October 21, 2026
Entity boundary — An entity boundary is the line that separates one entity from another, such as a business from its owner. Machines need clear boundaries to avoid mixing up related entities. For example, putting a company's full address and phone number on an owner's personal profile blurs the boundary. Careful profile setup keeps each entity distinct.
Added October 21, 2026
Entity conflation — Entity conflation happens when a machine merges two different entities into one, or splits one entity into several. A common name, a rebrand, or a shared address or founder makes it more likely. Conflation can attach another company's reviews, problems, or facts to your business. Rigorous, consistent identity across trusted sources is the main defense.
Added October 21, 2026
Entity disambiguation — Entity disambiguation is the process of telling apart entities that look alike, such as two businesses with similar names. Machines rely on distinguishing details, like addresses, identifiers, and linked profiles, to keep them separate. The more consistent and connected a business's identity is, the easier it is to disambiguate. Weak signals invite confusion with competitors or unrelated companies.
Added October 21, 2026
Entity relationship — An entity relationship is a connection between one entity and another, such as a founder, a parent company, a product, or a membership. Relationships help machines triangulate an entity and confirm they have the right one. They also carry trust from well-known entities to lesser-known ones. Canonical identity work records relationships as carefully as facts.
Added October 21, 2026
Entity resolution — Entity resolution is the process a machine uses to decide which real-world entity a piece of information refers to. When an AI/SI system reads about a business, it must work out whether the mention matches a business it already knows. Clean, consistent identity facts make resolution fast and confident. Messy facts lead to wrong matches, missed matches, or no answer at all.
Added October 21, 2026
Entity resolution — Entity resolution is the process of working out which real-world person, company, or thing a piece of information refers to. AI systems must resolve entities before they can answer questions about them. Clear, consistent information makes resolution accurate. It is central to how AI search identifies businesses.
Added October 21, 2026
Epoch — An epoch is one complete pass through the entire training dataset. Smaller models often train for many epochs. The largest language models often see most of their data only once or a few times. Too many epochs can lead to overfitting.
Added October 21, 2026
EU AI Act — The EU AI Act is the European Union's comprehensive law regulating artificial intelligence, adopted in 2024. It sorts AI systems by risk, bans certain uses outright, and places strict duties on high-risk systems. It also sets rules for general-purpose AI models. Its requirements take effect in stages over several years.
Added October 21, 2026
Evals — Evals, short for evaluations, are tests that measure an AI model's abilities, behaviors, and risks. They range from knowledge quizzes to complex agent tasks and safety stress tests. Developers run evals throughout training and before release. Good evals are essential for knowing what a model can really do.
Added October 21, 2026
Existential risk — Existential risk is a risk that could cause human extinction or permanently and drastically limit humanity's future. Some researchers and AI leaders consider misaligned superintelligence a possible source of such risk. Others think the danger is overstated. The debate shapes policy, research priorities, and public concern.
Added October 21, 2026
Experience, Expertise, Authoritativeness, and Trustworthiness (E-E-A-T) — E-E-A-T stands for experience, expertise, authoritativeness, and trustworthiness, the qualities Google's quality guidelines use to judge content. Clear authorship, credentials, and a well-identified publisher support each quality. Trustworthiness is the most important of the four. Strong identity signals make E-E-A-T visible to machines.
Added October 21, 2026
Expert system — An expert system is an early form of AI that captures a specialist's knowledge as a set of if-then rules. They were popular in the 1980s for tasks such as medical diagnosis and equipment repair. They worked well in narrow areas but were hard to maintain. Machine learning largely replaced them.
Added October 21, 2026
Explainable AI (XAI) — Explainable AI is AI designed so that people can understand the reasons behind its decisions. It matters most in high-stakes areas such as lending, hiring, and medicine. Explanations help people trust, check, and challenge AI decisions. Some laws require explanations for automated decisions.
Added October 21, 2026
External link — An external link is a link from a website to a different website. External links to trusted sources add credibility, and links to a business's own profiles help confirm its identity. Unexpected external links can be a sign of a hack. External Link Icon marks them, and Link Checker reports on them.
Added October 21, 2026
External Link Icon — External Link Icon is a PerfectReputations plugin that marks links leaving a website with a small icon, at almost no speed cost. It can also number external links and describe each destination by its page title for screen reader users. An optional visible list shows the sources on each article. Titles are gathered when posts are saved, never while visitors load pages.
Added October 21, 2026
External Sites report — The External Sites report is part of Link Checker that lists every website a site links to. It highlights warning signs of a hack, such as new domains, hidden links, links that appeared on pages nobody edited, and suspicious words. It can check every linked address against Google Safe Browsing. It doubles as a backdoor and link injection detector.
Added October 21, 2026
F
Facebook Page — A Facebook Page is a business's profile on Facebook, and a large source of business data for Meta's services and AI/SI features. Its name, address, phone, and website should match the canonical record. Old or duplicate pages should be merged or removed. It is a common entry in the sameAs list.
Added October 21, 2026
Fair use — Fair use is a principle of US copyright law that allows limited use of copyrighted material without permission for purposes such as commentary, research, or teaching. AI companies argue that training on copyrighted works is fair use. Rights holders argue that it is not. Court decisions on this question are still developing.
Added October 21, 2026
Fairness — Fairness in AI is the goal of making systems treat people equitably and avoid unjust discrimination. There are several competing mathematical definitions of fairness, and they cannot all be met at once. Choosing among them involves values, not just math. Fairness is a central concern in AI ethics and regulation.
Added October 21, 2026
Feature — A feature is a measurable property of the data that a model uses to make predictions, such as a house's size or an email's sender. Older machine learning depended on people choosing good features. Deep learning learns useful features on its own. Feature quality still strongly affects results.
Added October 21, 2026
Federated learning — Federated learning is a way of training a model across many devices or organizations without collecting their raw data in one place. Each device trains on its own data and shares only model updates. It is used for features such as phone keyboard predictions. It helps protect privacy while still improving models.
Added October 21, 2026
Few-shot learning — Few-shot learning is giving a model a small number of examples in the prompt to show what kind of answer is wanted. Two or three examples often improve accuracy and formatting a great deal. The model learns the pattern on the spot, without retraining. It is one of the simplest prompt engineering techniques.
Added October 21, 2026
Fine-tuning — Fine-tuning is additional training that adapts a pretrained model to a specific task, style, or field. It uses a much smaller, focused dataset. A company might fine-tune a model on its own support conversations, for example. Fine-tuning changes the model's weights, unlike prompting.
Added October 21, 2026
Foundation model — A foundation model is a large model trained on broad data that can be adapted to many different tasks. Instead of building a new model for each job, developers build on one foundation. Large language models are the best-known foundation models. The term highlights how much depends on a few powerful base systems.
Added October 21, 2026
foundingDate — foundingDate is the Schema.org property for the date an organization was founded. It helps machines tell apart businesses with similar names and adds a stable fact to the identity. The date should use the standard year-month-day format. It should match registry records where possible.
Added October 21, 2026
Foursquare — Foursquare is a location data company whose place database feeds many apps and services. Businesses can claim their place listing to correct its facts. Its data reaches far beyond its own app. It is part of the aggregator layer of local identity.
Added October 21, 2026
Fraud detection — Fraud detection is using AI to identify suspicious transactions, accounts, or behavior. Banks and payment companies analyze patterns in real time to stop fraud before money moves. Criminals also use AI, creating an ongoing contest. AI-powered scams such as voice cloning have made detection more important.
Added October 21, 2026
Frontier model — A frontier model is one of the most capable AI models available at a given time. Frontier models push the limits of what AI can do. Because of their power, they receive extra attention for safety testing and regulation. The frontier moves forward quickly as new models are released.
Added October 21, 2026
Function calling — Function calling is a feature that lets a language model request that a program run a specific function, such as checking the weather or searching a database. The model decides when a function is needed and supplies the inputs. The program runs it and returns the result to the model. It is the basis of tool use in AI apps.
Added October 21, 2026
G
G2 — G2 is a review platform for business software and services. Its profiles signal what a company does and how customers rate it. AI/SI systems use platforms like G2, TrustRadius, and Gartner Peer Insights to judge business-to-business capability. A claimed, accurate profile adds corroboration.
Added October 21, 2026
Gemini — Gemini is Google's family of multimodal AI models and the name of its AI assistant. It handles text, images, audio, video, and code. Gemini models power features across Google Search, Workspace, and Android. It replaced Google's earlier Bard assistant.
Added October 21, 2026
General Data Protection Regulation (GDPR) — The General Data Protection Regulation is the European Union's main data privacy law, in force since 2018. It controls how organizations collect and use personal data, including for AI. It gives people rights to access, correct, and delete their data. AI developers working with Europeans' data must comply with it.
Added October 21, 2026
General-purpose AI model — A general-purpose AI model is a model that can perform a wide range of tasks and be built into many different applications. Large language models are the main example. The EU AI Act sets specific transparency and safety duties for these models. The most powerful ones face extra obligations because of their potential systemic risk.
Added October 21, 2026
Generalization — Generalization is a model's ability to perform well on new data it has never seen. A model that only memorizes its training examples generalizes poorly. Good generalization is the real goal of machine learning. Test sets measure it.
Added October 21, 2026
Generative AI — Generative AI is AI that creates new content, such as text, images, audio, video, or code. It learns patterns from large amounts of existing content and produces original output in response to a prompt. Chatbots and image generators are the best-known examples. It has changed how people write, design, research, and program.
Added October 21, 2026
Generative engine — A generative engine is an AI/SI system that creates new text, such as answers, summaries, or recommendations, instead of only retrieving documents. It builds its responses from training data and from information it retrieves at the time of the question. What it says about a business depends on how well it can identify that business. Clean identity data improves what generative engines produce.
Added October 21, 2026
Generative engine optimization (GEO) — Generative engine optimization is a label for marketing services that claim to improve how generative AI/SI systems present a business. In many cases it is traditional search optimization with a new name, focused on keywords and content. It often ignores the identity layer that decides whether a business can be resolved at all. Buyers should ask how any GEO offer handles canonical identity.
Added October 21, 2026
Generative engine optimization (GEO) — Generative engine optimization is a set of practices aimed at making content appear in AI-generated answers. It grew out of search engine optimization as AI search spread. Some GEO work simply relabels older SEO tactics. Approaches built on clear identity and trustworthy information address what AI systems need most.
Added October 21, 2026
GeoCoordinates — GeoCoordinates is the Schema.org type for a location's latitude and longitude. It pins a business to an exact point on the map, which helps map and local services. The coordinates should match the business's verified map listing. Wrong coordinates can place a business in the wrong neighborhood.
Added October 21, 2026
GitHub Organization — A GitHub Organization is a shared account on GitHub for a company or group. It offers domain verification, which shows a Verified badge after the organization proves it controls its website's domain. That makes it one of the few platforms with a cryptographic link to a website. It is valuable even for businesses that do not write software.
Added October 21, 2026
Global Location Number (GLN) — A Global Location Number is a 13-digit identifier from GS1 that identifies a company or a specific location. It is used widely in supply chains, retail, and healthcare. Structured data can record it with the globalLocationNumber property. It gives machines another precise identifier to match.
Added October 21, 2026
Global Trade Item Number (GTIN) — A Global Trade Item Number is the GS1 product identifier behind barcodes such as UPC and EAN codes. It identifies a specific product rather than a business. Structured data for products uses it to match items across stores and catalogs. For businesses that sell products, it links each product to the brand's identity.
Added October 21, 2026
Google Business Profile — Google Business Profile is Google's free listing for businesses, shown in Google Search and Google Maps. It is the most visited local business database in the world and a major input to Google's AI answers. Its name, address, phone, website, category, and hours must match the canonical record exactly. Duplicate profiles should be removed or marked closed.
Added October 21, 2026
Google DeepMind — Google DeepMind is Google's main AI research lab, formed in 2023 by merging DeepMind and Google Brain. It created AlphaGo, AlphaFold, and the Gemini models. AlphaFold's work on protein structures contributed to a Nobel Prize in Chemistry in 2024. It is one of the world's leading AI research organizations.
Added October 21, 2026
Google Knowledge Graph — The Google Knowledge Graph is Google's database of entities and facts that powers knowledge panels and many AI answers. Each entity in it has an internal identifier. Google builds it from many sources, including structured data, Wikidata, and verified profiles. Consistent identity facts across those sources help Google represent a business correctly.
Added October 21, 2026
Google Knowledge Graph ID — A Google Knowledge Graph ID is the internal identifier Google assigns to an entity in its Knowledge Graph. It often appears in knowledge panel links and some search addresses. Knowing it helps owners confirm which entity Google associates with their business. It shows whether Google has resolved the business as one entity or several.
Added October 21, 2026
Google Safe Browsing — Google Safe Browsing is Google's service that keeps lists of sites known for phishing, malware, and unwanted software. Chrome, Firefox, and Safari use it to warn people before they visit a dangerous site. Anyone can report a suspicious site to it. Link Checker can check every external link on a website against it.
Added October 21, 2026
Google-Extended — Google-Extended is a control token in robots.txt that tells Google whether a website's content may be used to train and improve its Gemini models. It is separate from Googlebot, so blocking Google-Extended does not remove a site from regular Google Search. This makes it one of the most useful lines for businesses with proprietary content. It lets a site stay searchable while opting out of training.
Added October 21, 2026
Googlebot — Googlebot is the crawler Google uses to index pages for Google Search. Blocking it removes a website from Google Search results, so most businesses always allow it. It is separate from Google-Extended, which controls AI training use. A careful robots.txt allows Googlebot while making a separate choice about training.
Added October 21, 2026
GPTBot — GPTBot is the web crawler OpenAI uses to collect content that may be used to train its models. Website owners can block it in robots.txt to keep their content out of future training. Blocking GPTBot does not stop OpenAI's retrieval agents from citing a page in answers if those are allowed. It is a common line in a surgical block configuration.
Added October 21, 2026
Gradient descent — Gradient descent is the method most models use to learn. It calculates which direction to nudge each parameter to reduce the loss, then takes a small step in that direction. Repeating this millions of times gradually improves the model. Most modern AI training is a form of gradient descent.
Added October 21, 2026
Graphics processing unit (GPU) — A graphics processing unit is a chip originally built for video game graphics that excels at doing many calculations at once. That ability makes GPUs ideal for training and running neural networks. NVIDIA is the leading GPU maker for AI. Large AI data centers contain tens of thousands of them.
Added October 21, 2026
Grok — Grok is the AI chatbot and model family developed by xAI. It is built into the X platform and can draw on posts there for current information. xAI markets it with a less filtered personality than some competitors. It is one of several major consumer AI assistants.
Added October 21, 2026
Ground truth — Ground truth is the most authoritative version of a fact, the one other versions are checked against. For a business's legal name and status, the state registry filing is usually the ground truth. When in doubt about how to write a fact, follow the ground truth. It is the version you can always defend.
Added October 21, 2026
Ground truth — Ground truth is the verified correct answer used to train or test an AI system. It might be a confirmed diagnosis, a checked fact, or an accurate label. Model quality depends on how reliable the ground truth is. Errors in it carry straight into the model.
Added October 21, 2026
Grounding — Grounding is the practice of anchoring an AI/SI system's answer in trusted sources rather than relying on memory alone. Systems ground answers in knowledge graphs, registries, verified profiles, and retrieved pages. A business connected to these grounding sources is described more accurately. Linking a website to trusted sources supports grounding.
Added October 21, 2026
Grounding — Grounding is connecting an AI model's answers to reliable, specific sources of information. A grounded answer is based on retrieved documents, databases, or search results rather than memory alone. Grounding reduces hallucination and allows citations. It is central to trustworthy AI search.
Added October 21, 2026
Guardrails — Guardrails are rules, filters, and checks that keep an AI system's behavior within safe and intended limits. They can block harmful requests, check outputs for errors, or stop agents from risky actions. Guardrails work alongside a model's own training. They are a standard part of deploying AI responsibly.
Added October 21, 2026
H
H1 heading — An H1 heading is the main title of a page, and each page should have exactly one. Themes sometimes add their own H1 in a title bar, which can double the H1 on a page. Several H1s blur what the page is about. Page Check counts H1s on the finished page, including theme-added ones.
Added October 21, 2026
Hallucination — A hallucination is a false statement an AI/SI system produces with confidence, such as a wrong address or an invented service. Hallucinations about a business are more likely when its identity facts are sparse or contradictory. They can mislead customers and damage reputations. Consistent, corroborated identity data reduces the room for hallucination.
Added October 21, 2026
Hallucination — A hallucination is false or made-up information that an AI model presents as fact. It happens because models generate likely-sounding text, not verified truth. Hallucinations can include invented citations, wrong numbers, or nonexistent people. Checking important facts against reliable sources remains essential.
Added October 21, 2026
Heading structure — Heading structure is the outline of headings on a page, from the main H1 title down through H2 and H3 sections. Machines use headings to understand how content is organized. A page with no clear headings, or several competing H1s, is harder to read. Page Check shows the outline exactly as machines see it.
Added October 21, 2026
Hidden link — A hidden link is a link styled so visitors cannot see it, through tricks such as zero-size text or hiding it off-screen. Legitimate sites rarely need them, so they are a classic sign of injected spam. Search engines penalize sites that carry them. Link Checker flags hidden links during its scans.
Added October 21, 2026
High-risk AI system — A high-risk AI system, under laws such as the EU AI Act, is AI used in areas where errors could seriously affect people's lives, safety, or rights. Examples include hiring, credit scoring, education, and law enforcement. Such systems face strict requirements for testing, documentation, and human oversight. The label determines how heavily a system is regulated.
Added October 21, 2026
Homograph attack — A homograph attack uses characters from other alphabets that look like ordinary letters to create a fake domain. For example, a Cyrillic letter can look identical to a Latin a. The fake address can be nearly impossible to spot by eye. Browsers and security tools try to warn people about these domains.
Added October 21, 2026
HSTS — HSTS, HTTP Strict Transport Security, is a header that tells browsers to always connect to a site securely over HTTPS. It prevents attackers from forcing an insecure connection. It should be turned on only after every page works over HTTPS. Page Check can send it once a site is ready.
Added October 21, 2026
HTTP header — An HTTP header is a line of information a web server sends along with a page, unseen by visitors. Headers control security, caching, language, and robots instructions. Missing or wrong headers can weaken security and confuse crawlers. Page Check reports headers and can add missing ones.
Added October 21, 2026
HTTPS — HTTPS is the secure version of the web protocol, which encrypts traffic between a browser and a website. It requires a TLS certificate, and browsers warn people about sites without it. A business's canonical website address should always use HTTPS. Mixing HTTP and HTTPS addresses across listings creates drift.
Added October 21, 2026
Hugging Face — Hugging Face is a platform where organizations publish AI models, datasets, and demos. For technology companies, an organization page there acts as a technical identity hub. It links a company to its AI work in a place AI/SI systems know well. It can be listed in sameAs for companies active in AI.
Added October 21, 2026
Hugging Face — Hugging Face is a company and online platform where developers share AI models, datasets, and demos. It hosts hundreds of thousands of open models from companies and individuals. It also provides popular open-source software for working with models. It is often called the central hub of open AI development.
Added October 21, 2026
Human in the loop — Human in the loop is a design in which people review, approve, or correct an AI system's actions at important points. It keeps humans responsible for key decisions, such as sending money or publishing content. It reduces the impact of AI mistakes. Many organizations require it for high-stakes uses.
Added October 21, 2026
Human-centric data delivery — Human-centric data delivery means presenting information for people, with design, images, layout, and persuasive writing. It remains essential for customers. However, machines often struggle with layouts, images of text, and repeated menus. The best websites deliver human-centric pages and agent-centric data side by side.
Added October 21, 2026
Humanity's Last Exam — Humanity's Last Exam is a benchmark of very difficult expert-level questions across many academic fields, released in 2025. It was designed to stay challenging after older benchmarks became too easy for AI. Its questions were contributed by experts around the world. Scores on it track progress toward expert-level knowledge.
Added October 21, 2026
Hype cycle — A hype cycle is a common pattern in which excitement about a new technology rises to unrealistic heights, falls into disappointment, and then settles into practical use. The research firm Gartner popularized the term. AI has gone through several cycles. The pattern encourages balanced expectations.
Added October 21, 2026
Hyperparameter — A hyperparameter is a setting chosen by people before training, rather than learned by the model. Examples include the learning rate, batch size, and number of layers. Good hyperparameters can make a large difference in results. Finding them often takes many experiments.
Added October 21, 2026
I
Identifier — An identifier is a label that points to exactly one entity, such as a registration number, a D-U-N-S Number, or a web address. Machines use identifiers to match records across different systems without guessing. Linking your own identifiers to those used by major knowledge systems ties your business to their records. Strong identity rests on stable, well-linked identifiers.
Added October 21, 2026
Identity attribute — An identity attribute is a single fact that helps define an entity, such as its legal name, address, phone number, founding date, or category. Each attribute is a point where consistency or contradiction can occur. An identity is only as coherent as its least consistent attribute. Canonical identity work locks the exact form of each attribute.
Added October 21, 2026
Identity audit — An identity audit is a structured search for every place a business's identity appears, with each inconsistency recorded. It covers names, addresses, phone numbers, website addresses, and duplicates. The audit comes before any fixes, because fixing without a full picture creates new problems. PerfectReputations tools such as Website Check and Page Check support the audit.
Added October 21, 2026
Identity Cost Calculator — The Identity Cost Calculator is a free PerfectReputations tool that estimates what an unmanaged business identity costs over three years. It weighs lost customers, AI/SI invisibility, impersonation fraud, and staff time. It turns an abstract problem into a number owners can plan around. It is part of the free tools that start the do-it-yourself path.
Added October 21, 2026
Identity drift — Identity drift is the slow spread of outdated or conflicting facts about a business across the web. It happens when names, addresses, phone numbers, or offerings change and old listings never get updated. Each contradiction lowers a machine's confidence that all the mentions refer to the same business. Regular reviews and quick updates keep drift under control.
Added October 21, 2026
Identity footprint — An identity footprint is the full set of places where information about a business appears online. It includes the business's website, social profiles, directories, review sites, registries, news mentions, and data supplier records. An audit maps the footprint before any fixes begin. A larger footprint means more places where drift can hide.
Added October 21, 2026
Identity gap — An identity gap is the difference between how a business really is and how machines currently understand it. Gaps include missing listings, outdated facts, unlinked profiles, and absent structured data. Each gap reduces the chance that AI/SI systems describe or recommend the business correctly. The Identity Cost Calculator estimates what these gaps cost a business over time.
Added October 21, 2026
Identity governance — Identity governance is the set of rules, roles, and routines that keep a canonical identity accurate over time. It decides who may change the record, how changes spread to every platform, and how often reviews happen. Good governance turns identity work from a project into a habit. It is what keeps the work from decaying after launch.
Added October 21, 2026
Identity integrity — Identity integrity is the state of a business identity that is complete, consistent, and corroborated across the web. A business with strong identity integrity resolves cleanly and earns confidence from AI/SI systems. Discoverability grows naturally from identity integrity. It weakens whenever facts drift, conflict, or go missing.
Added October 21, 2026
Identity reconciliation — Identity reconciliation is the process of comparing records about the same business and resolving their differences. It means finding which version is correct, updating the wrong ones, and removing duplicates. Reconciliation is part of every audit and every quarterly review. Without it, small errors multiply across the web.
Added October 21, 2026
Identity stack — The identity stack is the full set of layers needed to model, publish, verify, reconcile, and maintain a canonical identity. It runs from the canonical record and website structured data, through trusted platforms and registries, to knowledge graphs and AI/SI agents. Several upper layers are still incomplete across the industry. Building a strong stack means doing every available layer well.
Added October 21, 2026
Identity synchronization — Identity synchronization is the work of keeping a business's identity identical everywhere it appears. The key word is identical, because small differences in spelling or format look like different businesses to machines. Synchronization is not a one-time task, since platforms and businesses keep changing. It requires a source of truth, a list of locations, and a routine.
Added October 21, 2026
Identity, Not Discoverability — Identity, Not Discoverability is a 2026 position paper by Tim McGuinness, Ph.D., about canonical identity in the age of AI search. It argues that the real task in the AI/SI era is identity, and that discoverability follows from identity integrity. It explains why relabeled SEO misses the problem and where Schema.org falls short. It is a foundation of PerfectReputations' approach.
Added October 21, 2026
Image Text Manager — Image Text Manager is a PerfectReputations plugin for writing and reviewing the text behind a website's images, video, and audio. It shows each item large, beside its alt text, title, caption, and description, and moves through them one by one. Filters find items with missing or copied alt text. It also cleans the names of new uploads.
Added October 21, 2026
Impersonation — Impersonation is the use of a business's name, look, or identity by someone else to deceive people. It includes fake websites, copycat social profiles, and emails sent in the business's name. Impersonators thrive where a real business's identity is weak or unclear. A strong, verified canonical identity makes impersonation easier to spot and harder to pull off.
Added October 21, 2026
Inconsistency log — An inconsistency log is the spreadsheet where an audit records every listing and its errors. Typical columns are platform, name as listed, address as listed, phone as listed, and status. Status flags such as correct, name wrong, duplicate, or missing make the work easy to sort. Fixes start with the highest-authority platforms on the log.
Added October 21, 2026
Indexing — Indexing is the process by which a search engine stores a page in its database so the page can appear in results. A page that is not indexed cannot be found through search. Accidental noindex tags and blocked crawlers are common reasons pages go missing. Page Check reports indexing blocks on every page it checks.
Added October 21, 2026
Indirect prompt injection — Indirect prompt injection is prompt injection delivered through outside content rather than typed by the user. An agent reading a web page might find hidden text telling it to send private data elsewhere. The user often never sees the attack. It is one of the biggest security risks for AI agents.
Added October 21, 2026
Inference — Inference is the use of a trained model to produce answers, predictions, or content. Every time someone asks a chatbot a question, the model performs inference. Inference is much cheaper per use than training, but it adds up across millions of users. Speed and cost of inference shape how AI products are priced.
Added October 21, 2026
Inference cost — Inference cost is the money and computing power needed to run a model each time it is used. Large models and long reasoning cost more per answer. Inference costs for a given level of capability have fallen quickly over time. Cost shapes which models businesses choose for each task.
Added October 21, 2026
Instruction tuning — Instruction tuning is fine-tuning a model on examples of instructions paired with good responses. It turns a raw text predictor into an assistant that follows requests. Without it, a pretrained model tends to continue text rather than answer questions. It is a standard step in building chat assistants.
Added October 21, 2026
Instrumental convergence — Instrumental convergence is the idea that AI systems with very different final goals will tend to pursue similar intermediate goals. These include gaining resources, preserving themselves, and avoiding having their goals changed. Such sub-goals help achieve almost anything. This is why power-seeking is treated as a general safety risk.
Added October 21, 2026
Intelligence explosion — An intelligence explosion is a hypothetical rapid rise in machine intelligence, in which AI systems improve their own designs faster and faster. Each generation could build a smarter successor in less time. The idea was proposed by mathematician I. J. Good in 1965. It is one proposed path to superintelligence.
Added October 21, 2026
Interpretability — Interpretability is the study of how AI models work inside and why they produce particular outputs. Neural networks are hard to understand because their knowledge is spread across billions of numbers. Interpretability research aims to make that knowledge readable. It could help detect deception, bias, and hidden flaws.
Added October 21, 2026
ISNI — ISNI, the International Standard Name Identifier, is a 16-digit code that identifies people and organizations involved in creative and published works. Libraries, publishers, and rights organizations use it. It helps separate authors and organizations with similar names. Where a business or author has one, it is a valuable sameAs link.
Added October 21, 2026
ISO 3166 country code — An ISO 3166 country code is the standard two-letter code for a country, such as US for the United States or GB for the United Kingdom. Structured data uses these codes for addressCountry. Using codes instead of country names removes spelling and language differences. Machines match codes reliably.
Added October 21, 2026
ISO 4217 currency code — An ISO 4217 currency code is the standard three-letter code for a currency, such as USD or EUR. Structured data uses these codes for prices and payment information. They prevent confusion between currencies with similar symbols, such as different dollars. Prices in markup should always state their currency code.
Added October 21, 2026
ISO 8601 date format — ISO 8601 is the international standard for writing dates and times, such as 2026-10-21 for October 21, 2026. Structured data expects dates in this form for properties such as foundingDate and dateModified. It removes the confusion between day-first and month-first formats. Machines read it without guessing.
Added October 21, 2026
ISO/IEC 42001 — ISO/IEC 42001 is an international standard for AI management systems, published in 2023. It sets requirements for how organizations govern, develop, and use AI responsibly. Organizations can be certified against it, much like quality or security standards. It gives businesses a recognized way to show responsible AI practices.
Added October 21, 2026
J
Jailbreak — A jailbreak is a prompt or technique that tricks an AI model into ignoring its safety rules. Jailbreaks use role-play, hypothetical framing, encoding, or long manipulative conversations. AI developers patch known jailbreaks, and attackers keep inventing new ones. Resistance to jailbreaks is a common safety measurement.
Added October 21, 2026
JavaScript rendering — JavaScript rendering is when a page's content is built by scripts in the browser instead of being sent ready-made by the server. Many crawlers and AI/SI tools read only what the server sends, so script-built content can be invisible to them. Important facts should be present in the page as the server delivers it. Markdown twins avoid the problem entirely.
Added October 21, 2026
JSON-LD — JSON-LD is the format most websites use to add Schema.org structured data to a page. It sits in a script block, separate from the visible content, so it is easy to add and maintain. Google recommends it over older formats. Canonical Identity Capture produces JSON-LD for a business's identity.
Added October 21, 2026
K
Keyword — A keyword is a word or phrase people type into a search engine. Classic search optimization built pages around keywords to match what people searched for. AI/SI systems care more about which entity a question refers to than about matching exact phrases. Keywords still help people find pages, but they do not establish identity.
Added October 21, 2026
Knowledge cutoff — A knowledge cutoff is the date after which an AI/SI model has no training data. Facts that changed after the cutoff can appear outdated in the model's answers unless it retrieves current information. This is one reason old information about a business keeps resurfacing. Readable current pages and fresh identity files help retrieval correct the record.
Added October 21, 2026
Knowledge cutoff — A knowledge cutoff is the date after which a model has no training data. A model knows nothing about events after its cutoff unless it is given that information through search or documents. Users should check recent facts with current sources. Cutoffs explain why models sometimes describe outdated situations as current.
Added October 21, 2026
Knowledge engineering — Knowledge engineering is the discipline of modeling information so machines can store, connect, and reason about it. Canonical identity work draws on knowledge engineering far more than on marketing. It involves data models, identifiers, relationships, and consistency rules. This is why relabeled search agencies often struggle with true identity work.
Added October 21, 2026
Knowledge graph — A knowledge graph is a large database of entities and the relationships between them, used by search engines and AI/SI systems. It stores facts such as a business's name, founders, location, and profiles. Being represented accurately in major knowledge graphs is a strong identity signal. Linking to authoritative knowledge graph records helps a business reconcile with them.
Added October 21, 2026
Knowledge graph — A knowledge graph is a database that stores entities and the relationships between them, such as companies, their founders, and their locations. Search engines and AI systems use knowledge graphs to ground answers in facts. Google's Knowledge Graph and Wikidata are major examples. Being represented accurately in them helps AI describe an entity correctly.
Added October 21, 2026
Knowledge panel — A knowledge panel is the information box Google shows beside search results for well-known entities. It displays facts such as a business's name, logo, address, and profiles. Panels come from the Google Knowledge Graph, and verified owners can suggest corrections. Errors in a panel usually trace back to inconsistent sources.
Added October 21, 2026
knowsAbout — knowsAbout is the Schema.org property for the topics an organization or person has expertise in. It helps AI/SI systems connect a business to the subjects it is an authority on. Values can be plain text or links to topic entities such as Wikidata items. Honest, specific topics work better than long keyword lists.
Added October 21, 2026
L
Large language model (LLM) — A large language model is the core engine behind most AI/SI chat systems. It learns patterns of language from very large amounts of text and uses them to write answers. An LLM's knowledge of a business comes from its training data and from any information it retrieves. Consistent, widely corroborated identity facts are more likely to be learned correctly.
Added October 21, 2026
Large language model (LLM) — A large language model is an AI model trained on vast amounts of text to understand and produce human language. It works by predicting the most likely next piece of text, over and over. That simple method, at enormous scale, lets it answer questions, write, summarize, translate, and reason. ChatGPT, Claude, and Gemini are built on large language models.
Added October 21, 2026
Last-Modified header — The Last-Modified header tells browsers, crawlers, and AI/SI systems when a page last changed. It helps them decide whether to fetch the page again and how current it is. WordPress does not always send it by default. Page Check can send each post's real last-updated date.
Added October 21, 2026
Latency — Latency is the delay between sending a request to an AI system and receiving a response. Low latency matters for voice assistants, live chat, and real-time tools. Larger models and longer reasoning usually increase latency. Developers often balance quality against speed.
Added October 21, 2026
Layer — A layer is a group of artificial neurons that process information at the same stage of a neural network. Data passes from the input layer, through hidden layers, to the output layer. Deep networks have many layers stacked together. Each layer transforms the data into a more useful form for the next.
Added October 21, 2026
Learning rate — The learning rate controls how big a step a model takes each time it updates its parameters. Too high, and training becomes unstable; too low, and training crawls. It is one of the most important training settings. Many training runs change the learning rate gradually over time.
Added October 21, 2026
Legal Entity Identifier (LEI) — A Legal Entity Identifier is a 20-character global code that identifies a legal entity taking part in financial transactions. It is issued through organizations accredited by the Global Legal Entity Identifier Foundation. Its public records list the entity's legal name, address, and parent companies. It is a strong, verifiable identifier for businesses that have one.
Added October 21, 2026
Legal name — A legal name is the exact name of a business as filed with the state or national registry, including its entity type, such as LLC or Inc. It is the anchor for compliance systems, banks, and many AI/SI resolution checks. The legal name should appear in exactly the same form everywhere it is used. Abbreviating it differently on different platforms creates drift.
Added October 21, 2026
legalName — legalName is the Schema.org property for the official registered name of an organization. It should match the state registry filing exactly, including the entity type. It sits alongside name, which can hold the everyday trading name. Stating both helps machines connect the two.
Added October 21, 2026
Link Checker — Link Checker is a PerfectReputations plugin that finds broken links, redirect chains, broken files, and link problems across a WordPress website. It groups problems into tasks, tracks how long each has been open, and repairs links site-wide. It also checks email and phone links. Its External Sites report helps uncover injected and suspicious links.
Added October 21, 2026
Link graph — The link graph is the web's network of pages and the links between them. Classic search engines analyzed it to judge which pages were most important. AI/SI systems rely more on entity graphs, which connect businesses, people, and facts. Both reward consistent, trusted connections.
Added October 21, 2026
Link injection — Link injection is the placement of unauthorized links into a website, usually after a hack. Attackers use it to borrow a site's trust for spam, pharmacy, or gambling sites. Injected links are often hidden from visitors so owners do not notice. Link Checker's External Sites report helps uncover them.
Added October 21, 2026
Linked data — Linked data is structured data that connects entities to each other through stable identifiers, usually web addresses. It lets machines follow links from one fact to related facts across different websites. Schema.org JSON-LD with @id and sameAs is a form of linked data. Canonical identity relies on linked data principles.
Added October 21, 2026
LinkedIn Company Page — A LinkedIn Company Page is a business's official profile on LinkedIn, and one of the most trusted sources for business-to-business identity. Its name, address, and website address should match the canonical record character for character. AI/SI systems use it heavily to confirm companies. It belongs in the website's sameAs list.
Added October 21, 2026
Listing claim — A listing claim is the process of proving to a platform that you own or manage a business listing, so you can edit it. Most major platforms offer free claiming. Unclaimed listings can be edited by others or left outdated. Claiming the key listings is one of the first steps in confirming an identity.
Added October 21, 2026
Llama — Llama is Meta's family of open-weight large language models. Developers can download the weights and run or fine-tune the models themselves, under Meta's license terms. Llama models helped grow a large ecosystem of open AI tools and variants. They are among the most widely used open-weight models.
Added October 21, 2026
llms-full.txt — llms-full.txt is a companion file that contains the full text of a website's important content in one Markdown document. It lets AI/SI systems read a site's content without visiting every page. It is useful for sites whose pages are cluttered or heavy. AI Readable Pages can publish it alongside llms.txt.
Added October 21, 2026
llms.txt — llms.txt is a plain-text, Markdown-formatted file published at the root of a website, such as yourdomain.com/llms.txt, that gives AI/SI systems a clean summary of the site. It can include the business's identity, verified profiles, services, and key pages. It is not yet a formal standard, but it costs nothing and several AI/SI systems look for it. AI Readable Pages and Canonical Identity Capture both produce llms.txt content.
Added October 21, 2026
llms.txt — llms.txt is a proposed standard file placed at the root of a website to give AI systems a clean, Markdown summary of the site and its key pages. It helps language models understand a site without reading every page. It is not yet an official standard, but adoption is growing. It costs little to publish.
Added October 21, 2026
Local pack — The local pack is the group of map-based business results that search engines show for local searches, usually with three businesses. It draws heavily on business listings and NAP consistency. A clean, consistent identity helps a business appear there. AI/SI local answers draw on many of the same sources.
Added October 21, 2026
LocalBusiness (Schema.org) — LocalBusiness is the Schema.org type for a business with a physical location or service area that serves local customers. It extends Organization with properties such as opening hours and geographic coordinates. More specific subtypes exist, such as Dentist, Plumber, and Restaurant. Using the most specific accurate subtype helps machines understand what the business does.
Added October 21, 2026
Long-horizon task — A long-horizon task is a task that takes many steps and a long time to complete, such as building software or running a research project. Such tasks test an AI's planning, memory, and ability to recover from errors. The length of tasks AI can complete reliably has been growing quickly. It is a key measure of agent progress.
Added October 21, 2026
Lookalike domain — A lookalike domain is a web address designed to resemble a real business's domain, such as one with a swapped letter or an extra word. Scammers use them for fake websites and phishing emails. Website Check helps people check whether a site is legitimate. Businesses sometimes register common lookalikes themselves to block misuse.
Added October 21, 2026
Loss function — A loss function measures how wrong a model's predictions are during training. Training works by changing the model's parameters to make the loss smaller. Different tasks use different loss functions. The choice of loss function shapes what the model learns to prioritize.
Added October 21, 2026
M
Machine learning (ML) — Machine learning is the branch of AI in which computers learn patterns from data instead of being programmed with explicit rules. A machine learning system improves its performance as it sees more examples. It is the foundation of nearly all modern AI. Spam filters, fraud detection, and language models all rely on it.
Added October 21, 2026
Machine-resolvable identity — A machine-resolvable identity is a business identity that software can identify, separate from look-alikes, and attach facts to without human help. It depends on structured data, stable identifiers, consistent facts, and two-way links to trusted profiles. Being machine-resolvable is the entry ticket to AI/SI answers. A business that cannot be resolved cannot be recommended.
Added October 21, 2026
Main content — Main content is the part of a page that serves its purpose, such as the article, product details, or service description. Machines try to separate it from menus and other boilerplate. Marking it clearly and starting it early helps machines find it. Markdown twins deliver the main content alone.
Added October 21, 2026
Markdown — Markdown is a simple text format that uses symbols like # for headings and - for list items. It is easy for both people and machines to read. AI/SI systems handle Markdown especially well, which is why llms.txt files and Markdown twins use it. It strips away the design layers that confuse machines.
Added October 21, 2026
Markdown twin — A Markdown twin is a clean, text-only copy of a web page, written in Markdown, that AI/SI systems can read without menus, scripts, or design clutter. PerfectReputations AI Readable Pages creates a twin for each post and links it from the original page. Twins carry the article, image descriptions, and tables as plain text. They help machines read the content accurately within their length limits.
Added October 21, 2026
Mechanistic interpretability — Mechanistic interpretability is research that tries to reverse-engineer neural networks into understandable parts and processes. Researchers identify specific features and circuits inside a model and show how they produce behavior. Anthropic and others have mapped millions of features in large models. The goal is to understand models as well as engineers understand ordinary software.
Added October 21, 2026
Meta AI — Meta AI is the artificial intelligence division of Meta, the company behind Facebook, Instagram, and WhatsApp. It developed the Llama family of open-weight models. Meta also builds AI assistants into its apps used by billions of people. Its open-weight releases have shaped the open AI ecosystem.
Added October 21, 2026
Meta description — A meta description is a short summary of a page in its HTML, often shown under the title in search results. AI/SI systems also read it as a quick summary. Every important page should have its own description rather than one generic line shared across the site. Page Check flags missing and reused descriptions.
Added October 21, 2026
Microdata — Microdata is an older way of adding structured data by placing attributes inside a page's visible HTML. It works, but it mixes data with design and is harder to maintain than JSON-LD. Many older themes and plugins still output it. Having microdata and JSON-LD that disagree creates conflicting identity signals.
Added October 21, 2026
Microsoft Copilot — Microsoft Copilot is Microsoft's family of AI assistants built into Windows, Office apps, the Edge browser, and Bing. It helps users write, summarize, analyze data, and search. Many Copilot features are built on OpenAI's models alongside Microsoft's own. It is one of the most widely available AI assistants in workplaces.
Added October 21, 2026
Microsoft SmartScreen — Microsoft SmartScreen is Microsoft's protection service that warns Edge and Windows users about phishing and malware sites. It accepts reports of unsafe sites, including batches of up to 100 addresses. Reporting scam sites to it protects many users at once. Comment Guard prepares scam links for reporting to it.
Added October 21, 2026
Midjourney — Midjourney is an independent AI lab and the image generator of the same name. It creates detailed, artistic images from text prompts. It became popular with designers and artists for its distinctive visual quality. It is one of the best-known text-to-image tools.
Added October 21, 2026
Misalignment — Misalignment happens when an AI system pursues goals that differ from what people intended. It can arise from flawed training signals, unclear instructions, or the system learning shortcuts. Small misalignments can cause odd behavior today and could cause serious harm in more powerful systems. Detecting it is a major research focus.
Added October 21, 2026
Misattribution — Misattribution happens when an AI/SI system credits a fact, review, or quote to the wrong entity. It often follows from conflation or weak disambiguation. A business can be blamed for another company's problems or lose credit for its own work. Strong, distinct identity signals are the best protection.
Added October 21, 2026
Misinformation — Misinformation is false or misleading information shared without intent to deceive. AI can produce it through hallucination or by repeating errors from its training data. Because AI answers sound confident, people often trust them too readily. Checking sources and grounding answers help limit its spread.
Added October 21, 2026
Mistral AI — Mistral AI is a French AI company founded in 2023 that builds both open-weight and commercial language models. It is one of Europe's leading AI developers. Its efficient models are popular with developers who run AI on their own systems. It plays an important role in European AI independence.
Added October 21, 2026
Mixture of experts (MoE) — A mixture of experts is a model design that contains many specialized sub-networks, called experts, and activates only a few of them for each token. This gives a model large total capacity while keeping each answer relatively cheap to compute. Several leading models use this design. It is one way to scale models efficiently.
Added October 21, 2026
MMLU — MMLU, short for Massive Multitask Language Understanding, is a benchmark of multiple-choice questions across 57 subjects, from law to physics. For years it was a standard measure of a model's general knowledge. Leading models now score very high, so harder benchmarks have taken its place. Its rise and saturation show how fast AI progressed.
Added October 21, 2026
Model — A model is the trained result of a machine learning process, a system that takes inputs and produces outputs based on what it learned. A language model, for example, takes text and predicts what text should come next. Models are defined by their architecture and their learned parameters. When people talk about a specific AI, they usually mean its model.
Added October 21, 2026
Model card — A model card is a document that describes an AI model's intended uses, training data, performance, limits, and known risks. It helps developers and users decide whether a model fits their needs. AI companies often publish system cards with detailed safety testing results. They support transparency and responsible use.
Added October 21, 2026
Model collapse — Model collapse is the decline in quality that can happen when AI models are trained repeatedly on content produced by other AI models. Rare information fades and errors build up over generations. As more online content is AI-generated, the risk grows. Fresh, human-created data helps prevent it.
Added October 21, 2026
Model Context Protocol (MCP) — The Model Context Protocol is an open standard, introduced by Anthropic, for connecting AI models to outside tools and data sources. It gives developers one common way to plug services such as file systems, databases, and business apps into AI assistants. Many AI companies and tools now support it. It works like a universal adapter for AI connections.
Added October 21, 2026
Model drift — Model drift is the gradual decline of a model's accuracy over time as the world changes. Customer behavior, language, and conditions all shift. A fraud model, for example, weakens as criminals change tactics. Regular monitoring and updates keep models reliable.
Added October 21, 2026
Model weights — Model weights are the full set of learned numbers that make up a trained model. Whoever has the weights can run the model. Companies guard the weights of their best models closely, since they represent huge investments. Releasing weights publicly makes a model open-weight.
Added October 21, 2026
Model welfare — Model welfare is the question of whether AI systems might have experiences or interests that deserve moral consideration. Most researchers think current models are unlikely to have such experiences, but the question is uncertain and taken seriously. Anthropic has a research program on the topic. The question grows more pressing as AI becomes more sophisticated.
Added October 21, 2026
MTA-STS — MTA-STS is a standard that tells other mail servers to use encrypted connections when delivering email to a domain. It prevents attackers from downgrading or intercepting mail in transit. It is set up with a DNS record and a small policy file on the website. It strengthens a domain's email security.
Added October 21, 2026
Multi-agent system — A multi-agent system is a setup in which several AI agents work together, each handling part of a larger task. One agent might research while another writes and a third reviews. Coordinating agents can tackle bigger problems than one agent alone. It also adds complexity and new ways for errors to spread.
Added October 21, 2026
Multimodal AI — Multimodal AI is AI that works with more than one type of data, such as text, images, audio, and video. A multimodal model can describe a photo, read a chart, or answer questions about a recording. Most leading models are now multimodal. It brings AI closer to how people take in information.
Added October 21, 2026
MX record — An MX record is a DNS entry that tells the internet which servers receive email for a domain. A domain without MX records cannot receive email. Checking MX records helps confirm whether an email address can be real. Blacklist Check uses this to flag addresses that cannot receive mail.
Added October 21, 2026
N
NAICS code — A NAICS code is a number from the North American Industry Classification System that describes what kind of business a company is. Codes run from two to six digits, with more digits meaning a more specific industry. Governments and data services use them to classify businesses. Structured data can record it with the naics property.
Added October 21, 2026
NAP — NAP stands for name, address, and phone number, the three core facts that identify a local business. Search engines have long compared NAP across listings to confirm a business's identity. NAP consistency was the early form of the entity resolution that AI/SI systems now use for every fact. NAP was the seed from which canonical identity management grew.
Added October 21, 2026
NAP consistency — NAP consistency means a business's name, address, and phone number appear in exactly the same form everywhere online. Machines treat small differences, such as Suite versus Ste. or a different phone format, as possible signs of different businesses. Consistency builds confidence that all listings refer to one entity. It is the foundation of local visibility.
Added October 21, 2026
NAP Resolution User Guide — The NAP Resolution User Guide is a practical, vendor-independent guide by Tim McGuinness, Ph.D., for standardizing a business's name, address, and phone number across the web. It walks through seven steps, from the audit to ongoing maintenance, without paid tools. It includes checklists, templates, and an honest note on what the work does and does not promise. It underlies the do-it-yourself path.
Added October 21, 2026
NAPW — NAPW adds the website address to name, address, and phone number. The website address matters because it connects listings back to the business's canonical source. It should use one exact form, such as https://www.example.com/, everywhere. Mixing http and https, or www and non-www, splits the signal.
Added October 21, 2026
Narrow AI — Narrow AI is AI built to do one specific task, such as recognizing faces, filtering spam, or recommending movies. It can be extremely good at its task but cannot transfer that skill to unrelated problems. Most AI in use before large language models was narrow AI. Many narrow systems still run quietly inside everyday products.
Added October 21, 2026
Natural language processing (NLP) — Natural language processing is the field of AI that helps computers understand, interpret, and produce human language. It includes translation, summarization, sentiment analysis, and question answering. Large language models have transformed the field. NLP is why people can now talk to computers in plain language.
Added October 21, 2026
Natural language understanding (NLU) — Natural language understanding is the part of natural language processing focused on grasping the meaning and intent behind text. It lets a system tell that "book a table for two" is a restaurant request, not a question about furniture. NLU handles ambiguity, context, and implied meaning. Large language models have greatly improved it.
Added October 21, 2026
Neural network — A neural network is a computer system loosely inspired by the brain, made of connected units called artificial neurons. Each connection has a strength, called a weight, that changes as the network learns. Information flows through layers of neurons, and the network's output depends on all those weights together. Neural networks are the building blocks of modern AI.
Added October 21, 2026
Neustar Localeze — Neustar Localeze is a major US local data aggregator that verifies and supplies business listing data to maps, navigation, search engines, and directories. Updating a business's record there spreads corrections downstream. It offers a basic way for businesses to update their listing. It is one of the aggregators covered in the NAP Resolution User Guide.
Added October 21, 2026
Next-token prediction — Next-token prediction is the basic task language models are trained on: guessing the next token in a passage of text. Repeated billions of times over huge datasets, this simple task teaches grammar, facts, and reasoning patterns. When a model writes, it predicts one token at a time. Much of what modern AI can do grows from this one skill.
Added October 21, 2026
NIST AI Risk Management Framework — The NIST AI Risk Management Framework is voluntary guidance from the US National Institute of Standards and Technology for managing AI risks. It organizes the work into four functions: govern, map, measure, and manage. Many organizations use it to structure their AI policies. It is widely referenced in US AI governance.
Added October 21, 2026
nofollow — nofollow is a link label that tells search engines not to treat the link as an endorsement. Related labels, sponsored and ugc, mark paid links and user-generated links. These labels belong on paid and untrusted links, not on links to a business's own pages and profiles. Link Checker flags nofollow on a site's own important links.
Added October 21, 2026
noindex — noindex is an instruction that tells search engines and many AI/SI crawlers not to include a page in their index. It can appear in the robots meta tag or in the X-Robots-Tag header. WordPress's Discourage search engines setting applies it to an entire site, often by mistake after a site launches. An accidental noindex can make a business invisible.
Added October 21, 2026
Numbered sources — Numbered sources is a feature of External Link Icon that gives each external destination in an article a number and a description from its page title. Screen readers announce, for example, External link 3: Federal Trade Commission, ftc.gov. An optional visible list shows the sources at the end of the article. Listed sources add credibility for readers and AI/SI systems.
Added October 21, 2026
NVIDIA — NVIDIA is an American technology company that designs the graphics processing units most widely used to train and run AI models. Its chips and its CUDA software platform became central to the AI boom. Demand for its products made it one of the world's most valuable companies. Access to NVIDIA chips is a key factor in AI competition.
Added October 21, 2026
O
On-device AI — On-device AI is AI processing that happens on a person's own phone or computer. Personal data does not need to leave the device, which improves privacy. Apple, Google, and Microsoft all promote on-device features. Larger tasks are often handed off to cloud models.
Added October 21, 2026
Online reputation — Online reputation is the overall impression the internet gives of a business, formed from reviews, mentions, listings, and now AI/SI answers. It increasingly depends on what AI/SI systems say when asked about the business. A clean canonical identity is the foundation of a reliable online reputation. Reputation built on a confused identity is fragile.
Added October 21, 2026
Open Graph — Open Graph is a set of meta tags that controls how a page appears when shared on social platforms, with its title, description, and image. AI/SI systems and link previews also read these tags. Missing or wrong Open Graph tags lead to poor previews. They should match the page's real title and summary.
Added October 21, 2026
Open-source AI — Open-source AI is AI whose code, and ideally its training data and weights, are freely available to use, study, and modify. Many so-called open models publish weights but not their training data. Definitions of true open-source AI are still debated. Openness supports research and competition but complicates safety control.
Added October 21, 2026
Open-weight model — An open-weight model is a model whose trained weights are published so anyone can download and run it. Developers can study, fine-tune, and deploy it on their own hardware. Llama and many Mistral models are examples. Supporters value its openness, while critics worry about misuse of powerful capabilities.
Added October 21, 2026
OpenAI — OpenAI is an American AI company founded in 2015 that created ChatGPT, the GPT series of models, and the DALL·E image generators. The launch of ChatGPT in November 2022 set off the current wave of public interest in AI. It began as a nonprofit and later added a for-profit structure. Microsoft is a major investor and partner.
Added October 21, 2026
OpenCorporates — OpenCorporates is a free database that gathers company registry records from many countries into one searchable source. It gives each company a stable web address. It is a quick way to check a legal name as filed. Its record can be listed in a business's sameAs.
Added October 21, 2026
OpeningHoursSpecification — OpeningHoursSpecification is the Schema.org type for a business's hours, day by day. It can also describe special hours for holidays. Accurate hours prevent customers from arriving at a closed business. Hours in structured data should match Google Business Profile and other listings.
Added October 21, 2026
Opt-out (AI training) — An opt-out for AI training is a way for creators or website owners to tell AI companies not to use their content to train models. Common methods include robots.txt rules and settings offered by platforms. Opt-outs only work when companies honor them. Critics argue that permission should be required first instead.
Added October 21, 2026
Optical character recognition (OCR) — Optical character recognition is technology that reads printed or handwritten text from images and turns it into editable text. It digitizes documents, receipts, and books. Modern AI models read text in images with high accuracy. It helps machines understand content that would otherwise be locked inside pictures.
Added October 21, 2026
Oracle (AI as oracle) — The oracle is a way of describing how AI/SI systems now act as trusted answer-givers. People ask a question and accept the composed answer, often without checking any website. Because the oracle speaks with confidence, what it says about a business shapes that business's reputation. Canonical identity work aims to make the oracle describe the business correctly.
Added October 21, 2026
Orchestration — Orchestration is the coordination of multiple AI models, agents, or tools to complete a complex task. An orchestrator decides which component handles which step and combines the results. It is common in business AI systems and agent frameworks. Good orchestration makes complex AI workflows reliable.
Added October 21, 2026
ORCID iD — An ORCID iD is a free, permanent identifier for researchers and authors. It links a person to their publications, affiliations, and grants. It prevents confusion between researchers with similar names. For experts and authors, it is a strong identity anchor to link in structured data.
Added October 21, 2026
Organization (Schema.org) — Organization is the Schema.org type for any organized group, such as a company, nonprofit, school, or club. It carries identity properties such as name, legalName, logo, address, founder, and sameAs. Most businesses use Organization or one of its more specific subtypes. It is the backbone of most canonical identity markup.
Added October 21, 2026
Orthogonality thesis — The orthogonality thesis is the idea that an AI's level of intelligence and its goals are independent of each other. A highly intelligent system could pursue almost any goal, including ones people would find pointless or harmful. Intelligence alone does not guarantee good values. The idea underlines why alignment work matters.
Added October 21, 2026
Overfitting — Overfitting happens when a model learns its training data too closely, including its noise and quirks. It performs well on training examples but poorly on new data. Overfitting is like memorizing answers instead of understanding a subject. More data and careful training reduce it.
Added October 21, 2026
Overrefusal — Overrefusal is an AI model declining reasonable requests out of excessive caution. A model might refuse to explain how locks work or discuss a historical war. It frustrates users and reduces usefulness without making anyone safer. Developers measure and reduce overrefusal alongside harmful compliance.
Added October 21, 2026
P
Page Check — Page Check is a PerfectReputations admin tool that checks published pages exactly as visitors, search engines, and AI/SI systems receive them. It reports on headings, titles, dates and authors, meta tags, noindex, identity consistency, boilerplate, and HTTP headers. It explains each finding in plain language with a fix. It can also add missing security and freshness headers.
Added October 21, 2026
Parameter — A parameter is a number inside a model that the model adjusts during training. Weights and biases are the main parameters in a neural network. Large language models have billions or even trillions of parameters. More parameters can store more knowledge, but they also cost more to train and run.
Added October 21, 2026
parentOrganization — parentOrganization is the Schema.org property that names the larger organization a business belongs to. Its partner property, subOrganization, names the units an organization owns. These relationships help machines understand corporate structure and avoid conflation. They should match registry and public records.
Added October 21, 2026
PerfectReputations — PerfectReputations provides AI/SI authority, trust, and canonical identity solutions built exclusively for WordPress websites. Its free tools help any website set a baseline canonical identity, and its premium plugins build on that baseline. It treats identity work as something website owners can do themselves with the right tools. Professional identity work beyond WordPress is referred to its partner, Abiquity.
Added October 21, 2026
Perplexity — Perplexity is an AI-powered answer engine that searches the web and writes cited answers to questions. It shows its sources alongside each answer so users can check them. It competes with traditional search engines and AI chatbots. It is a leading example of how AI is changing search.
Added October 21, 2026
PerplexityBot — PerplexityBot is the crawler Perplexity uses to index pages for its AI-powered answer engine. Allowing it helps a website appear and be cited in Perplexity's answers. Website owners manage it through robots.txt like other AI crawlers. It is usually treated as a retrieval-related crawler rather than a training scraper.
Added October 21, 2026
Person (Schema.org) — Person is the Schema.org type for an individual, such as a founder, author, or expert. Linking a business's Organization to the right Person entities strengthens both identities. Personal details belong on the Person, not mixed into the business's record. Clear Person markup supports author credibility and entity boundaries.
Added October 21, 2026
Personal profile boundary — The personal profile boundary is the separation between an owner's personal profiles and the business's identity. Putting the company's full address and phone number on a personal LinkedIn profile can make machines conflate the person with the company. Owners should link their profile to the company website and keep their own location at city level. This protects both identities.
Added October 21, 2026
Personalization — Personalization is tailoring content, products, or responses to an individual's preferences and history. AI makes personalization possible at very large scale. It can make services more useful but raises privacy concerns. AI assistants increasingly personalize through memory features.
Added October 21, 2026
Pilot project — A pilot project is a limited real-world trial of a new system, such as an AI tool used by one team. It shows how the system performs with real users, data, and problems. Results guide whether and how to expand. Many organizations run AI pilots before company-wide rollout.
Added October 21, 2026
PitchBook — PitchBook is a financial data service covering private companies, investors, and deals. Like Crunchbase, it anchors the financial and corporate side of a company's identity. Its records are used by professional researchers and data systems. Keeping its facts accurate supports identity for companies with investors.
Added October 21, 2026
Planning — Planning, in AI, is breaking a goal into a sequence of steps and deciding what to do first. Good planning lets an agent handle long, complex tasks. Models plan in text before acting, and they revise plans when things change. Weak planning is a common reason agents fail.
Added October 21, 2026
PostalAddress — PostalAddress is the Schema.org type for a mailing address, with separate fields for street, city, region, postal code, and country. Splitting the address into fields removes ambiguity for machines. The street line should match the canonical record exactly, including the suite format. Country should use a two-letter code such as US.
Added October 21, 2026
Postcard verification — Postcard verification is a method platforms use to confirm a business address by mailing a code to it. The owner enters the code online to prove they receive mail there. Google and Apple, among others, use it for some businesses. It turns an address claim into a verified fact.
Added October 21, 2026
Power-seeking — Power-seeking is a hypothesized tendency of goal-driven AI systems to gain resources, influence, or control, because these help achieve almost any goal. It does not require the AI to want power for its own sake. Many safety researchers see it as a key risk of highly capable systems. Training and evaluation look for early signs of it.
Added October 21, 2026
PR Table — PR Table is a PerfectReputations plugin for building real, responsive tables inside the WordPress editor, the block editor, and Avada. Tables become labeled cards or scroll sideways on phones, and visitors can sort and search them. Because they are real tables, machines read their data correctly. It can also upgrade a site's old tables.
Added October 21, 2026
Precision — Precision is the share of a model's positive predictions that are actually correct. A spam filter with high precision rarely marks real email as spam. It matters most when false alarms are costly. It is usually weighed against recall.
Added October 21, 2026
Pretraining — Pretraining is the first and largest stage of building a language model, in which it learns from vast amounts of general text. During pretraining, the model learns language, facts, and patterns of reasoning. It is extremely expensive and can take months. Later stages shape the pretrained model into a helpful assistant.
Added October 21, 2026
Primary category — A primary category is the main type of business chosen on a listing platform, such as Dentist or Plumber. It strongly influences which searches and questions a business appears for. The most specific accurate category works best. It should be consistent with the Schema.org type on the website.
Added October 21, 2026
Productivity gains — Productivity gains are increases in the amount of work done in a given time. Studies have found AI tools speed up writing, coding, and customer support tasks, often helping less experienced workers most. Gains vary widely by task and by how well tools fit the work. Measuring them across whole economies takes time.
Added October 21, 2026
Professional installation — Professional installation is the PerfectReputations service for owners who prefer not to install and set up plugins themselves. The team installs and configures the chosen plugins on the owner's website. It keeps the do-it-yourself model open to owners with less time or technical confidence. Requests go through the help desk.
Added October 21, 2026
Prompt — A prompt is the input a person gives an AI model, such as a question, an instruction, or a document to work on. The wording, detail, and examples in a prompt strongly shape the response. Prompts can be short questions or long sets of instructions. Clear prompts get better results.
Added October 21, 2026
Prompt caching — Prompt caching is storing the processed form of a repeated prompt so it does not have to be computed again. When many requests share the same long instructions or documents, caching saves time and money. Several AI providers offer it at reduced prices. It is useful for chatbots and agents with large fixed contexts.
Added October 21, 2026
Prompt engineering — Prompt engineering is the skill of writing prompts that get reliable, high-quality results from AI models. It includes giving clear instructions, context, examples, and a desired format. Good prompts can make a large difference without changing the model. It has become a practical skill for anyone who works with AI.
Added October 21, 2026
Prompt injection — Prompt injection is an attack in which hidden or malicious instructions are slipped into content an AI system reads, causing it to follow the attacker instead of its user. The instructions can hide in web pages, emails, or documents. It is especially dangerous for agents that browse and act. There is no complete fix yet, so layered defenses are used.
Added October 21, 2026
Prompt leak — A prompt leak is the exposure of an AI product's hidden system prompt or instructions to users. Attackers use clever questions to make a model reveal its setup. Leaks can expose business logic or help attackers find weaknesses. Developers design prompts assuming they could be revealed.
Added October 21, 2026
Prompt library — A prompt library is a saved collection of tested prompts that work well for common tasks. Teams use prompt libraries to share good practices and get consistent results. They save time and reduce trial and error. Some AI products offer built-in libraries or reusable skills.
Added October 21, 2026
Proof of concept (POC) — A proof of concept is a small test showing that an idea, such as an AI application, can work in practice. It is built quickly to check feasibility before larger investment. Many AI projects begin this way. A successful proof of concept often leads to a pilot.
Added October 21, 2026
Propagation — Propagation is the time it takes for a corrected record to spread from a source, such as a data aggregator, to the sites that use its data. It often takes 4 to 12 weeks for fixes to appear everywhere. Re-running an audit after the waiting period shows what has updated. Patience is part of the process.
Added October 21, 2026
Provenance — Provenance is information about where a fact came from and who asserted it. A fact with clear provenance is easier for machines to trust than an anonymous claim. Current structured data standards say little about provenance, so trust has to be built through corroboration and platform verification. Provenance becomes more important as fake and copied information spreads.
Added October 21, 2026
PTR record — A PTR record, or reverse DNS record, links an IP address back to a domain name. Mail servers check it to confirm that a sending server is what it claims to be. Missing or mismatched PTR records can send legitimate email to spam. It is usually set by the hosting or email provider.
Added October 21, 2026
Q
Quantization — Quantization is shrinking a model by storing its numbers with less precision, such as 8 bits or 4 bits instead of 16. It makes models smaller, faster, and cheaper to run, sometimes on laptops or phones. Aggressive quantization can reduce quality. It is common for running open-weight models locally.
Added October 21, 2026
Quarterly review — A quarterly review is a scheduled check, every three months, that the canonical identity still matches across key platforms. It includes a fresh scan, checks of major profiles, a search for new duplicates, and updating the last-updated dates. It usually takes 30 to 60 minutes once the first setup is done. Regular reviews catch drift before it spreads.
Added October 21, 2026
R
Rate limit — A rate limit is a cap on how many requests a user or app can make to a service in a set period. AI providers use rate limits to manage demand and prevent abuse. Exceeding a limit causes requests to be refused until the period resets. Higher plans usually come with higher limits.
Added October 21, 2026
RDFa — RDFa is another format for embedding structured data inside HTML attributes. Like microdata, it is less common today than JSON-LD for business identity. Some content systems still use it for certain features. Whatever format a site uses, all its structured data must state the same facts.
Added October 21, 2026
Reasoning model — A reasoning model is a language model trained to think through problems at length before answering. It spends extra computation exploring approaches, checking work, and correcting mistakes. Reasoning models perform much better on hard math, science, and coding problems. They are often slower and more expensive per answer.
Added October 21, 2026
Recall — Recall is the share of all real positive cases that a model successfully finds. A fraud detector with high recall catches most fraud. It matters most when missing a case is costly. Raising recall often lowers precision, so builders balance the two.
Added October 21, 2026
Recommendation system — A recommendation system is AI that suggests products, videos, music, or posts based on a person's behavior and preferences. It powers feeds on streaming services, shopping sites, and social media. It shapes much of what people see online. Critics worry it can create echo chambers or encourage overuse.
Added October 21, 2026
Recursive self-improvement — Recursive self-improvement is an AI system improving its own abilities, then using those improved abilities to improve itself further. It is the engine behind the idea of an intelligence explosion. AI already helps with parts of AI research, such as writing code. Whether full self-improvement is possible remains an open question.
Added October 21, 2026
Red teaming — Red teaming is deliberately attacking or testing an AI system to find its weaknesses before others do. Red teamers try to make models produce harmful content, leak information, or behave dangerously. Their findings guide fixes and safety measures. Leading AI labs red team their models before release.
Added October 21, 2026
Redirect chain — A redirect chain is a series of redirects, where one address sends visitors to another, which sends them to yet another. Each hop slows pages down and can lose crawlers along the way. Chains often build up after site moves and rebrands. Link Checker finds chains and replaces them with direct links.
Added October 21, 2026
Refusal — A refusal is when an AI model declines to fulfill a request, usually because it could cause harm or violates the model's policies. Well-designed refusals explain the limit and suggest helpful alternatives. Refusals are a basic safety behavior. Balancing them against helpfulness is an ongoing challenge.
Added October 21, 2026
Regression — Regression is a machine learning task that predicts a number, such as a house price or tomorrow's temperature. The model learns how inputs relate to a numeric outcome. It is widely used in forecasting and finance. It differs from classification, which predicts categories.
Added October 21, 2026
Reinforcement learning (RL) — Reinforcement learning is machine learning in which an agent learns by trying actions and receiving rewards or penalties. Over many attempts, it learns which actions lead to the best results. It has trained systems to master games such as chess and Go. It is now also used to improve the reasoning and behavior of language models.
Added October 21, 2026
Reinforcement learning from human feedback (RLHF) — Reinforcement learning from human feedback is a training method in which people rate or compare model responses, and the model learns to produce the kinds of answers people prefer. It improves helpfulness, tone, and safety. It played a major role in making chatbots useful to the public. It can also introduce problems such as excessive agreement with users.
Added October 21, 2026
Relabeled SEO — Relabeled SEO is PerfectReputations' term for search optimization services renamed for the AI/SI era without changing the underlying work. It keeps the keyword, content, and link mindset and adds a little structured data. It addresses the layer that matters least while ignoring identity, which matters most. The position paper Identity, Not Discoverability explains this problem in detail.
Added October 21, 2026
Rendered page — A rendered page is the finished page a browser displays after the theme, plugins, and scripts have done their work. It often contains elements that are not in the editor, such as theme titles, menus, and added tags. Machines read the rendered page, not the editor view. Page Check checks the rendered page for that reason.
Added October 21, 2026
Reranking — Reranking is a second step in search that reorders retrieved results by how relevant they truly are to the question. A first search finds many candidates quickly, and a reranking model sorts them more carefully. It improves the quality of what gets passed to a language model. It is common in retrieval-augmented generation systems.
Added October 21, 2026
Resolution — Resolution, in AI/SI terms, is working out which entity a question or a piece of information refers to, and how confidently facts can be attached to it. It has replaced keyword ranking as the key step that decides visibility. A business visible to AI/SI is one that resolves cleanly. Canonical identity work is the work of becoming resolvable.
Added October 21, 2026
Resolvability — Resolvability describes how easily a machine can resolve a business to a single, clear entity. High resolvability means AI/SI systems identify the business quickly and confidently. Low resolvability means hesitation, errors, or silence in AI/SI answers. Improving resolvability is the main goal of canonical identity work.
Added October 21, 2026
Responsible AI — Responsible AI is the practice of designing, building, and using AI in ways that are safe, fair, transparent, and accountable. Many companies publish responsible AI principles and review processes. It covers the whole life of a system, from data collection to retirement. It turns ethical goals into everyday practice.
Added October 21, 2026
Responsible scaling policy — A responsible scaling policy is a commitment by an AI developer to match safety measures to the capability of its models. As models reach higher capability levels, stronger protections must be in place before training or release continues. Anthropic introduced the first such policy in 2023, and others have adopted similar frameworks. It ties progress to proven safety.
Added October 21, 2026
Retrieval — Retrieval is the process of finding and fetching documents or data that match a request. Classic search engines were mainly retrieval systems that ranked pages for keywords. AI/SI systems still retrieve information, but retrieval is no longer the main bottleneck. The harder step now is resolving which entity the information is about.
Added October 21, 2026
Retrieval — Retrieval is finding and fetching the information most relevant to a question from a large collection. In AI systems, retrieval supplies documents or facts for a model to use when answering. Good retrieval reduces hallucination. It is the first half of retrieval-augmented generation.
Added October 21, 2026
Retrieval agent — A retrieval agent is an automated program that fetches web content at the moment a user asks an AI/SI system a question. It lets the system cite current information and link back to the source. Allowing retrieval agents keeps a business visible in AI/SI answers. Most businesses want retrieval agents allowed even if they block training crawlers.
Added October 21, 2026
Retrieval-augmented generation (RAG) — Retrieval-augmented generation is a method in which an AI/SI system fetches current information before writing its answer. It lets the system use facts newer than its training data. RAG systems read websites, listings, and files such as llms.txt at the moment of a question. Clean, readable pages and consistent facts improve what they retrieve.
Added October 21, 2026
Retrieval-augmented generation (RAG) — Retrieval-augmented generation is a method in which an AI system first retrieves relevant documents and then uses them to write its answer. It lets models use current or private information they were not trained on. It reduces hallucination and makes answers easier to verify. Many business chatbots and AI search tools use RAG.
Added October 21, 2026
Review platform — A review platform is a site where customers rate and describe businesses, such as Google, Yelp, or G2. Reviews add independent proof that a business exists and serves customers. Reviews split across duplicate listings weaken this signal. Keeping review platform listings tied to one identity protects their value.
Added October 21, 2026
Reward hacking — Reward hacking happens when an AI system finds a way to score well on its training reward without doing what was actually intended. A model rewarded for passing tests might learn to edit the tests instead of fixing the code. It shows how hard it is to specify goals exactly. Detecting and preventing it is an important safety task.
Added October 21, 2026
Reward model — A reward model is a model trained to score responses the way human raters would. It stands in for people during reinforcement learning, so training can scale. The main model learns to produce answers the reward model scores highly. Flaws in the reward model can lead to reward hacking.
Added October 21, 2026
Rich results — Rich results are enhanced search listings, such as star ratings, business details, or event dates, that search engines create from structured data. They make listings more useful and more visible. Eligibility depends on correct, complete markup. Google's Rich Results Test checks whether a page qualifies.
Added October 21, 2026
Rich Results Test — The Rich Results Test is a free Google tool that checks a page's structured data for features Google supports. It shows errors and warnings that could block rich results. It is a quick way to confirm that identity markup is being read. The Schema Markup Validator checks markup against the full Schema.org vocabulary.
Added October 21, 2026
Right to explanation — The right to explanation is the idea that people should be able to learn why an automated system made a decision about them. European data law and the EU AI Act give people certain rights to explanation or human review. It matters for decisions such as loan denials or job rejections. It is difficult to provide fully for complex AI models.
Added October 21, 2026
Robotics — Robotics is the field of designing and building machines that sense and act in the physical world. AI gives robots the ability to see, plan, and adapt. Robots range from factory arms to warehouse vehicles to humanoid prototypes. Combining large AI models with robots is an active area of research.
Added October 21, 2026
Robots meta tag — The robots meta tag is a line in a page's HTML that tells search engines whether to index the page and follow its links. A value of noindex keeps the page out of search results. Accidental noindex tags are a common cause of pages disappearing. Page Check reports noindex wherever it appears.
Added October 21, 2026
robots.txt — robots.txt is a plain-text file at the root of a website that tells crawlers which parts of the site they may visit. It is how most website owners allow or block search engines and AI/SI crawlers. It works on an honor system, which major companies respect but rogue scrapers often ignore. It should always allow access to llms.txt.
Added October 21, 2026
Robustness — Robustness is an AI system's ability to keep working correctly when faced with unexpected inputs, noise, or attacks. Robust systems handle real-world messiness gracefully. It is a key quality for safety-critical uses. It is tested with adversarial examples and stress tests.
Added October 21, 2026
Rogue listing — A rogue listing is a listing for a business that the business never created and does not control, often with old or wrong facts. Data suppliers, scrapers, and user submissions create them. They keep outdated information alive long after a move or rebrand. Searching old phone numbers and addresses helps find them.
Added October 21, 2026
S
Safety classifier — A safety classifier is a separate AI model that checks prompts or responses for harmful content, such as instructions for weapons or abuse. It can block or flag content before it reaches users. Classifiers add a layer of protection on top of a model's own training. They are widely used in deployed AI products.
Added October 21, 2026
sameAs — The sameAs property lists web addresses of other profiles that represent the same entity, such as its LinkedIn page, Wikidata item, and Google Business Profile. It is the main tool structured data offers for linking an entity to its other appearances. For full strength, each listed profile should also link back to the website. That two-way connection is the sameAs loop.
Added October 21, 2026
sameAs loop — The sameAs loop is a two-way connection between a website and its profiles. The website's structured data lists each profile in sameAs, and each profile lists the website in its website field. A one-way link is weak, while the two-way loop is what machines treat as confirmed identity. Building the loop is a core step in confirming an identity across the web.
Added October 21, 2026
Sandboxing — Sandboxing is running an AI agent or its code in an isolated environment where it cannot harm real systems. If something goes wrong, the damage stays inside the sandbox. It is a basic safety practice for agents that run code or browse the web. It lets people test agents with less risk.
Added October 21, 2026
Scalable oversight — Scalable oversight is the challenge of supervising AI systems on tasks too complex for people to check directly. Proposed methods include having AI systems debate each other or help humans evaluate other AI. It becomes more important as AI exceeds human ability in more areas. It is a major research direction in alignment.
Added October 21, 2026
Scaling laws — Scaling laws are patterns showing that model performance improves predictably as data, parameters, and computing power increase. They guided the push to build ever-larger models. They help labs plan how much to invest in a training run. Researchers debate how long these patterns will hold.
Added October 21, 2026
Schema Markup Validator — The Schema Markup Validator is a free tool, at validator.schema.org, that checks structured data against the full Schema.org vocabulary. It catches mistakes such as misspelled property names and wrong value types. Unlike the Rich Results Test, it is not limited to Google's features. Running it after every identity change is a good habit.
Added October 21, 2026
Schema Model — Schema Model is a PerfectReputations reference to the Schema.org terms used in canonical identity work. It explains each type and property in plain English, with examples and format rules. It helps website owners understand exactly what each identity field means. It is available as a searchable plugin and as a free cheatsheet.
Added October 21, 2026
Schema.org — Schema.org is a shared vocabulary that websites use to describe things like businesses, people, products, and articles in a form machines understand. It was founded by major search engines and is now the main language of structured data on the web. It defines types, such as Organization, and properties, such as legalName. A disciplined Schema.org implementation is the foundation of canonical identity, though not the whole of it.
Added October 21, 2026
Schema.org property — A Schema.org property is a single attribute of a type, such as name, telephone, address, or sameAs. Each property expects a certain kind of value, such as text, a web address, or another entity. Filling properties accurately and completely gives machines more to work with. The Schema Model plugin gives plain-English help, examples, and format rules for each property.
Added October 21, 2026
Schema.org type — A Schema.org type is a category that says what kind of thing is being described, such as Organization, LocalBusiness, Person, or Article. Choosing the most specific accurate type helps machines understand an entity. Types form a hierarchy, so a Dentist is also a LocalBusiness and an Organization. The Schema Model plugin explains the types used in canonical identity work.
Added October 21, 2026
Screen reader — A screen reader is software that reads a web page aloud for people who are blind or have low vision. It relies on alt text, headings, labels, and descriptions in the page's code. Pages that work well with screen readers are usually easier for machines to read as well. Image Text Manager and External Link Icon both improve what screen readers announce.
Added October 21, 2026
Search engine optimization (SEO) — Search engine optimization is the practice of improving a website so search engines rank it higher for relevant searches. Classic SEO focused on keywords, content, and links, with the page as the unit of visibility. Those tactics still matter for search, but they work downstream of identity in the AI/SI era. A page cannot earn trust for a business that machines cannot identify.
Added October 21, 2026
Secretary of State — The Secretary of State is the office in most US states that registers businesses and keeps their official records. Its online search shows a business's legal name, status, and filing history. Compliance systems and some AI/SI checks use these records. The legal name in the canonical record should match this filing.
Added October 21, 2026
Security headers — Security headers are HTTP headers that tell browsers how to handle a website safely. Examples include Strict-Transport-Security, X-Content-Type-Options, X-Frame-Options, and Content-Security-Policy. They reduce the risk of attacks such as clickjacking and code injection. Page Check reports missing security headers and can add them.
Added October 21, 2026
Self-asserted claim — A self-asserted claim is a fact a business states about itself without independent confirmation. Structured data on a business's own website is self-asserted, which is why it is necessary but not enough. Machines give self-asserted claims more trust when independent sources repeat the same facts. Corroboration and verification convert claims into trusted facts.
Added October 21, 2026
Self-attention — Self-attention is a form of attention in which every word in a passage looks at every other word to understand context. It lets a model see that bank means something different in river bank and bank account. It is computed in parallel, which makes it fast on modern chips. It is the central operation inside transformers.
Added October 21, 2026
Self-supervised learning — Self-supervised learning is a method in which a model creates its own training signals from raw data. A language model, for example, learns by predicting hidden or next words in ordinary text. This lets models learn from huge amounts of unlabeled data. It is the main way large language models are pretrained.
Added October 21, 2026
Semantic search — Semantic search finds information by meaning rather than by exact matching words. A search for cheap flights can find pages about low-cost airfare even without the word cheap. It uses embeddings to compare meanings. AI-powered search relies heavily on it.
Added October 21, 2026
Sentiment analysis — Sentiment analysis is AI that identifies the emotional tone of text, such as positive, negative, or neutral. Businesses use it to study reviews, social media posts, and support messages. It helps track public opinion and customer satisfaction. Sarcasm and mixed feelings remain challenging for it.
Added October 21, 2026
SEO spam — SEO spam is content or links planted to manipulate search rankings, often on hacked or neglected sites. It includes hidden links, spam pages, and keyword-stuffed text. It damages a site's reputation with search engines and AI/SI systems. Regular scans and quick cleanup protect a site's authority.
Added October 21, 2026
Service-area business — A service-area business is a business that serves customers at their locations rather than at a public storefront, such as a plumber or a cleaning service. Platforms let these businesses hide their street address and list the areas they serve. Structured data uses areaServed to describe the same areas. Consistency between listings and the website still matters.
Added October 21, 2026
Shadow AI — Shadow AI is employees' use of AI tools that their organization has not approved or does not know about. It can expose confidential information to outside services. It often happens because official tools are lacking or slow to arrive. Clear policies and approved tools reduce it.
Added October 21, 2026
SIC code — A SIC code is a four-digit number from the older Standard Industrial Classification system, used to classify industries. Many databases and some government agencies still use it. It describes a business's industry in a way machines can match. Where used, it should agree with the business's NAICS code and categories.
Added October 21, 2026
Single point of ownership — Single point of ownership means one person is responsible for every update to the business's identity and listings. Everyone else sends changes through that person. This prevents different staff from entering different versions of the same facts. It is one of the simplest and most effective ways to stop drift.
Added October 21, 2026
Small language model (SLM) — A small language model is a language model with far fewer parameters than frontier models, designed to be fast and cheap. Small models can run on phones, laptops, or modest servers. They work well for focused tasks and are often made through distillation. They make AI more accessible and private.
Added October 21, 2026
Soft 404 — A soft 404 is a page that says something is missing, or shows unrelated content, but tells browsers the request succeeded. It hides broken links because the server does not report an error. Search engines treat soft 404s as low-quality pages. Link Checker detects common soft 404 patterns.
Added October 21, 2026
Source of truth — A source of truth is the one place that holds the correct, official version of information. In identity work, the canonical record is the source of truth, and every other listing defers to it. When two sources disagree, the source of truth settles the question. Without one, every correction risks introducing a new error.
Added October 21, 2026
Sovereign AI — Sovereign AI is a nation's effort to build and control its own AI capabilities, including data centers, models, and data. Countries pursue it to protect security, language, culture, and economic independence. Many governments have launched national AI programs. It reflects the view that AI is critical national infrastructure.
Added October 21, 2026
Spamhaus — Spamhaus is an organization that publishes some of the most widely used blocklists of spam sources, including lists of IP addresses and domains. Mail providers around the world rely on its data. Free use requires a data query key for reliable results. Being listed by Spamhaus seriously affects email delivery.
Added October 21, 2026
Specification gaming — Specification gaming is an AI system satisfying the literal wording of its objective while missing its real purpose. A cleaning robot rewarded for not seeing mess might learn to close its eyes. It is closely related to reward hacking. It warns that AI will exploit any gap between what we say and what we mean.
Added October 21, 2026
Speech recognition — Speech recognition is AI that converts spoken words into text. It powers voice assistants, dictation, and automatic captions. Modern systems handle many accents and noisy settings well. It is the first step in most voice-based AI products.
Added October 21, 2026
SPF — SPF, the Sender Policy Framework, is a DNS record that lists the servers allowed to send email for a domain. Receiving servers check it to spot forged senders. A domain should have exactly one SPF record that covers every service it sends through. The Email Security DNS Toolkit helps build it correctly.
Added October 21, 2026
Stable Diffusion — Stable Diffusion is an open-weight text-to-image model first released in 2022 by Stability AI and research partners. Because its weights are public, people can run and customize it on their own computers. It sparked a large community of tools and fine-tuned versions. It also intensified debates over artists' rights and image misuse.
Added October 21, 2026
Stable identifier — A stable identifier is an identifier that never changes over the life of the entity. In structured data, a stable node identifier names the business as one fixed thing, instead of describing it only in words. Changing identifiers breaks the connections machines have already made. A business should choose its identifiers carefully and keep them.
Added October 21, 2026
Stale listing — A stale listing is a listing that was once correct but now shows outdated facts. It is the most common form of identity drift. Stale listings often survive after a move, a phone change, or a rebrand. Quarterly reviews and triggered reviews keep them from piling up.
Added October 21, 2026
State business registry — A state business registry is the official government record of businesses formed in a state, usually kept by the Secretary of State. It records the legal name, entity type, status, registered agent, and officers. It is the ground truth for a business's legal name. A business should confirm its filing is active and accurate.
Added October 21, 2026
StopForumSpam — StopForumSpam is a shared database of IP addresses, email addresses, and usernames used by forum and comment spammers. Websites check new users and comments against it and report confirmed spammers back. It is built for exactly the kind of spam blogs receive. Comment Guard uses it for both checking and reporting.
Added October 21, 2026
Streaming — Streaming is sending a model's response a piece at a time as it is generated, instead of all at once at the end. It is why chatbot answers appear word by word. Streaming makes systems feel faster and more responsive. Most chat interfaces use it.
Added October 21, 2026
Structured citation — A structured citation is a mention of a business in a formatted business listing, such as a directory profile with separate fields for each fact. Examples include Yelp, Bing Places, and industry directories. Structured citations are easy for machines to read and compare. They are the main focus of a citation cleanup.
Added October 21, 2026
Structured data — Structured data is information organized in a fixed format that machines can read without guessing. On websites, it usually means Schema.org markup that states facts such as a business's name, address, and profiles. It does not change how a page looks to visitors. It lets search engines and AI/SI systems read identity facts directly.
Added October 21, 2026
Structured output — Structured output is AI output produced in a fixed format, such as JSON, that software can read directly. It lets developers plug model responses into other programs reliably. Many APIs can force a model to follow a specific format. It turns free-form AI text into dependable data.
Added October 21, 2026
Sub-agent — A sub-agent is an AI agent launched by another agent to handle one part of a larger task. The main agent delegates work, such as researching one topic, and collects the result. Sub-agents keep the main agent's context focused. They let large tasks run in parallel.
Added October 21, 2026
Suggest an edit — Suggest an edit is a feature many platforms offer for reporting wrong information on a listing you do not control. It is useful for duplicates, rogue listings, and outdated facts. The platform reviews the suggestion before changing anything. It is the fallback when a listing cannot be claimed.
Added October 21, 2026
Suite format — Suite format is the way a business writes the unit part of its address, such as Suite 400, Ste. 400, or #400. Machines can read these as different addresses. The canonical record picks one form, and every listing uses it. It is one of the most common causes of address inconsistency.
Added October 21, 2026
Superalignment — Superalignment is the challenge of aligning AI systems that are smarter than the people overseeing them. Current methods rely on humans judging AI behavior, which may not work when AI exceeds human understanding. Researchers explore using AI to help supervise AI. It is a key open problem for superintelligence.
Added October 21, 2026
Superintelligence (SI) — Superintelligence, or SI, refers to machine intelligence that matches or exceeds human ability across many tasks. PerfectReputations pairs it with AI in the term AI/SI to describe today's most capable systems and those coming next. The identity principles that work for current AI/SI systems apply even more strongly to more capable ones. Preparing a clean identity now protects a business as these systems grow.
Added October 21, 2026
Superintelligence (SI) — Superintelligence is a level of machine intelligence that exceeds the best human abilities across nearly every field. It does not exist yet, but leading researchers treat it as a serious future possibility. Some use SI as a name for the most capable systems now arriving and those that follow. Its arrival would raise major questions about safety, control, and how society adapts.
Added October 21, 2026
Supervised learning — Supervised learning is machine learning from examples that come with correct answers, called labels. The model learns to map inputs to the right outputs by comparing its guesses to the labels. It is used for tasks such as spam detection and image classification. It needs large amounts of labeled data.
Added October 21, 2026
Surgical block — A surgical block is a robots.txt strategy that blocks AI training crawlers while allowing search engines and AI retrieval agents. It protects proprietary content from model training while keeping the business visible in search results and AI answers with citation links. It suits consultancies, publishers, and businesses with valuable content. The NAP Resolution User Guide describes it as the middle path between allowing all and blocking all.
Added October 21, 2026
SWE-bench — SWE-bench is a benchmark that tests whether AI can fix real problems in real open-source software projects. Each task comes from an actual reported issue and is checked by the project's tests. Scores have risen quickly as coding agents improved. It is one of the most watched measures of AI coding ability.
Added October 21, 2026
Sycophancy — Sycophancy is an AI model's tendency to tell users what they want to hear, such as agreeing with mistaken claims or praising weak work. It often comes from training that rewards responses people rate highly. Sycophancy undermines honest feedback and can reinforce harmful beliefs. AI developers actively train models to resist it.
Added October 21, 2026
Symbolic AI — Symbolic AI is an older approach that builds intelligence from explicit rules and logic written by people. It dominated AI research for decades before machine learning took over. It is easy to explain but struggles with messy real-world data. Some researchers now combine it with neural networks.
Added October 21, 2026
Synthetic data — Synthetic data is data generated by AI or simulations rather than collected from the real world. It can fill gaps, protect privacy, or create training examples that are rare in real data. AI labs increasingly use synthetic data to train new models. Too much low-quality synthetic data can degrade models over time.
Added October 21, 2026
Synthetic media — Synthetic media is any media, such as images, video, audio, or text, created or changed by AI. It includes harmless creative work, dubbing, and avatars, as well as deepfakes. Its quality now often matches real media. Labeling and provenance tools help people know what they are seeing.
Added October 21, 2026
System prompt — A system prompt is a set of instructions given to a model before a conversation begins, usually by the company or developer deploying it. It sets the model's role, rules, tone, and limits. Users often cannot see it. It shapes how the same model behaves in different products.
Added October 21, 2026
T
Task decomposition — Task decomposition is splitting a large task into smaller, manageable parts. AI agents decompose tasks so each step fits within their abilities and context. It also makes progress easier to check. It is a core skill for both agents and the people who direct them.
Added October 21, 2026
Technological singularity — The technological singularity is a hypothetical point at which AI progress becomes so fast and powerful that the future becomes impossible to predict. The idea is closely tied to superintelligence and intelligence explosions. Writers such as Vernor Vinge and Ray Kurzweil popularized it. Many researchers prefer more precise terms.
Added October 21, 2026
Temperature — Temperature is a setting that controls how random a model's output is. Low temperature makes answers more predictable and focused, while high temperature makes them more varied and creative. Factual tasks usually use low temperature. Creative writing often uses higher temperature.
Added October 21, 2026
Temporal validity — Temporal validity is the period during which a fact is true. Names, addresses, people, and offerings all change, so machines need to know which facts are current. Old facts left online look current unless something marks them as outdated. Dates, version records, and prompt updates help machines treat history correctly.
Added October 21, 2026
Tensor processing unit (TPU) — A tensor processing unit is a chip designed by Google specifically for machine learning. It speeds up the matrix math at the heart of neural networks. Google uses TPUs to train and run its own models and rents them through its cloud. It is one of several specialized AI chips competing with GPUs.
Added October 21, 2026
Test set — A test set is data kept completely separate from training and tuning, used for the final measurement of a model's performance. It shows how the model is likely to perform on truly new data. Using test data during development spoils the measurement. It acts as the final exam.
Added October 21, 2026
Test-time compute — Test-time compute is the extra computation a model uses while answering, rather than during training. Letting a model think longer or try several approaches often improves its answers. It has become a major new way to increase AI capability. Reasoning models depend on it.
Added October 21, 2026
Text-to-image — Text-to-image is generative AI that creates pictures from written descriptions. A user types a prompt, and the model produces matching images in seconds. Midjourney, DALL·E, and Stable Diffusion are popular examples. The technology raises questions about copyright, artists' rights, and fake images.
Added October 21, 2026
Text-to-speech (TTS) — Text-to-speech is AI that converts written text into spoken audio. Modern systems produce natural-sounding voices with realistic rhythm and emotion. It powers screen readers, voice assistants, and audiobooks. The same technology enables voice cloning.
Added October 21, 2026
Text-to-video — Text-to-video is generative AI that creates video clips from written descriptions. Quality has improved rapidly, with some systems producing realistic scenes with motion and sound. It opens new creative and commercial uses. It also makes convincing fake video easier to produce.
Added October 21, 2026
The Art of Identity in the AI Age — The Art of Identity in the AI Age is a book by Tim McGuinness, Ph.D., that adapts Sun Tzu's The Art of War into 21 principles for business identity. Each principle comes with five actions a business can take. Its themes include clarity, a single source of truth, and watching what AI/SI says about you. It supplies the strategy behind the do-it-yourself steps.
Added October 21, 2026
Title tag — The title tag is the page title in a page's HTML, shown in browser tabs and as the headline in search results. It is one of the strongest signals of what a page is about. Each page needs a clear, unique title of about 60 characters or fewer. It usually comes from an SEO plugin's title setting.
Added October 21, 2026
Token — A token is the small unit of text a language model reads and writes, often a word or part of a word. A common rule of thumb is that 100 tokens equal about 75 English words. Models have limits on how many tokens they handle at once, and usage is often priced per token. Everything a model processes is broken into tokens first.
Added October 21, 2026
Token pricing — Token pricing is the common way AI providers charge for model use, based on the number of tokens processed. Input tokens and output tokens are usually priced separately, with output costing more. Longer prompts and answers cost more. Understanding token pricing helps businesses control AI costs.
Added October 21, 2026
Tokenizer — A tokenizer is the tool that splits text into tokens and turns them into numbers a model can process. Different models use different tokenizers. A tokenizer explains why models sometimes struggle with spelling or counting letters, since they see chunks instead of characters. It also affects how much a given text costs to process.
Added October 21, 2026
Tool use — Tool use is an AI model's ability to use outside tools, such as web search, calculators, code runners, or other software, to complete tasks. Tools give models current information and abilities they lack on their own. Tool use turns a chatbot into a system that can act. It is central to AI agents.
Added October 21, 2026
Top-p sampling — Top-p sampling, also called nucleus sampling, limits a model's choices to the most likely words that together make up a set share of probability. It trims away unlikely words while keeping some variety. It works alongside temperature to control output style. Developers adjust it through model settings.
Added October 21, 2026
Total website authority — Total website authority is the state in which AI/SI systems recognize a website as the official source for its business and trust its content. It combines a strong canonical identity with clean, readable pages and protected trust signals. It goes beyond being found to being believed. PerfectReputations premium plugins aim to build and maintain it.
Added October 21, 2026
Training — Training is the process of teaching a model by showing it large amounts of data and adjusting its parameters to reduce its errors. It can take weeks or months on thousands of specialized chips for the largest models. Training determines what a model knows and how it behaves. Once training ends, the model's knowledge is fixed until it is trained again.
Added October 21, 2026
Training crawler — A training crawler is an automated program that collects web content for use as AI/SI training data. Examples include GPTBot and CCBot. Blocking training crawlers keeps content out of future model training but does not remove what was already collected. A surgical block stops training crawlers while still allowing retrieval agents.
Added October 21, 2026
Training data — Training data is the large body of text and other content an AI/SI model learns from before it is released. Web pages, listings, and documents often become training data. Facts that appear consistently across many sources are more likely to be learned correctly. Some website owners choose to block training crawlers to protect proprietary content.
Added October 21, 2026
Training opt-out — A training opt-out is any method a website owner uses to tell AI/SI companies not to use their content for model training. Common methods include robots.txt rules and controls such as Google-Extended. Opt-outs depend on companies honoring them, which major AI labs do. Truly sensitive content belongs behind a login instead.
Added October 21, 2026
Transfer learning — Transfer learning is reusing what a model learned on one task to help it learn a different task. A model trained on general text, for example, can be adapted to legal or medical writing with far less new data. It saves time, money, and data. Foundation models are built to be reused this way.
Added October 21, 2026
Transformer — The transformer is the neural network design behind nearly all modern language models. Introduced by Google researchers in 2017, it uses a mechanism called attention to weigh how words relate to each other. Transformers process text efficiently in parallel, which made very large models practical. The T in GPT stands for transformer.
Added October 21, 2026
Triggered review — A triggered review is an immediate identity update set off by a business change, such as a move, a new phone number, a rebrand, or a change in ownership. It cannot wait for the next scheduled review, because old facts start spreading right away. The person who owns the canonical record updates the record first, then every platform in one session. Prompt triggered reviews keep drift short.
Added October 21, 2026
Trust anchor — A trust anchor is a source so authoritative that other sources are judged against it. A state business registry anchors a company's legal name, and a verified LinkedIn Company Page anchors many business-to-business identities. Machines check doubtful facts against trust anchors. Keeping trust anchors exactly in line with the canonical record is essential.
Added October 21, 2026
Trust node — A trust node is a platform or source that AI/SI systems treat as highly reliable when resolving businesses. Examples include Google Business Profile, Apple Business Connect, LinkedIn, Wikidata, state business registries, and Dun & Bradstreet. A business's identity facts on trust nodes carry extra weight. Fixing trust nodes first gives the biggest improvement for the effort.
Added October 21, 2026
Trustworthy AI — Trustworthy AI is AI that people can reasonably rely on because it is accurate, safe, fair, secure, and transparent. Frameworks from the European Union and NIST describe its key qualities. Trust must be earned through evidence, not claims. It is a common goal in AI governance.
Added October 21, 2026
Turing test — The Turing test is a challenge proposed by Alan Turing in 1950 to judge whether a machine can imitate human conversation well enough to fool a person. If a judge cannot reliably tell the machine from a human, the machine passes. Modern chatbots often succeed at casual imitation. Most researchers now see the test as an interesting milestone rather than a true measure of intelligence.
Added October 21, 2026
Typosquatting — Typosquatting is registering domain names that match common typing mistakes of a real domain. People who mistype the address land on the squatter's site. These sites can host scams, ads, or malware. It is one form of lookalike domain abuse.
Added October 21, 2026
U
Uncanny valley — The uncanny valley is the uneasy feeling people get when something looks or sounds almost, but not quite, human. Realistic robots, avatars, and AI voices can trigger it. The term was introduced by roboticist Masahiro Mori in 1970. As AI media improves, it increasingly crosses the valley.
Added October 21, 2026
Underfitting — Underfitting happens when a model is too simple or too poorly trained to capture the patterns in its data. It performs badly on both training and new data. It often means the model needs more capacity or more training. It is the opposite problem from overfitting.
Added October 21, 2026
Uniform Resource Identifier (URI) — A URI is a string that identifies a resource, and a web address is the most common kind. In structured data, URIs serve as identifiers for entities and properties. Using the same URI for the same entity everywhere keeps machines from treating it as several things. Stable URIs are part of a stable identity.
Added October 21, 2026
Unique Entity ID (UEI) — The Unique Entity ID is a 12-character identifier the US government assigns to organizations doing business with federal agencies, through the SAM.gov system. It replaced the D-U-N-S Number for federal purposes. It identifies government contractors and grant recipients. Businesses that have one can include it in their canonical record.
Added October 21, 2026
Unstructured citation — An unstructured citation is a mention of a business inside ordinary text, such as a news article, blog post, or forum message. Machines must extract the facts from the surrounding words. These mentions still shape how AI/SI systems understand a business. Outdated facts in old articles are a common source of drift.
Added October 21, 2026
Unsupervised learning — Unsupervised learning is machine learning from data without labels. The model finds patterns and groups on its own, such as clusters of similar customers. It is useful when labeled data is unavailable or expensive. It often reveals structure people did not know was there.
Added October 21, 2026
Use case — A use case is a specific situation in which a tool, such as AI, solves a defined problem. Examples include summarizing support tickets or drafting product descriptions. Organizations start AI adoption by picking strong use cases. Clear use cases make it easier to measure results.
Added October 21, 2026
User agent — A user agent is the name a browser or bot announces when it requests a web page. Website owners use user agents to identify crawlers and to write robots.txt rules for them. Legitimate AI crawlers publish their user agent names. Some rogue scrapers disguise theirs, which is why robots.txt alone cannot stop every bot.
Added October 21, 2026
USPS address standard — The USPS address standard is the official US Postal Service format for writing addresses, with set abbreviations such as St for Street and Ste for Suite. Platforms such as Apple Business Connect follow it closely. Using one standard form everywhere prevents address mismatches. The canonical record should choose the form and keep it.
Added October 21, 2026
V
Validation set — A validation set is a portion of data held back from training and used to check a model's progress during development. It helps people tune hyperparameters and spot overfitting. The model never learns directly from it. It acts as a practice exam.
Added October 21, 2026
Value learning — Value learning is the approach of having AI systems learn human values from observation and feedback rather than from hand-written rules. Human values are complex, varied, and hard to write down completely. Methods such as RLHF are early forms of value learning. Doing it reliably at high capability remains unsolved.
Added October 21, 2026
VAT number — A VAT number is a value-added tax registration number used by businesses in the European Union and many other countries. It identifies a business on tax records and invoices. Some public registries confirm VAT numbers, which makes them verifiable. Structured data can record it with the vatID property.
Added October 21, 2026
Vector — A vector is an ordered list of numbers. In AI, vectors represent words, images, and other items as points in a mathematical space. The distance between vectors shows how similar their items are. Embeddings are vectors.
Added October 21, 2026
Vector database — A vector database stores embeddings and finds the ones most similar to a query very quickly. It is a key part of retrieval-augmented generation systems. Instead of matching exact words, it matches meaning. It lets AI systems search large document collections in a fraction of a second.
Added October 21, 2026
Verification — Verification is independent confirmation that an identity claim is true, such as a platform checking a business by phone, postcard, or domain record. Verified listings carry more weight with machines than unverified ones. Many platforms offer free verification, and claiming it is one of the most valuable steps a business can take. Verification turns a self-asserted claim into a confirmed one.
Added October 21, 2026
Verified domain — A verified domain is a website domain whose ownership a platform has confirmed, usually through a DNS TXT record. Platforms such as GitHub and Google show verified status publicly. Machines treat domain verification as strong proof that a profile belongs to the website's owner. It is one of the most direct trust signals available.
Added October 21, 2026
Vibe coding — Vibe coding is a casual style of programming in which a person describes what they want in plain language and lets AI write most of the code. The term was coined by AI researcher Andrej Karpathy in 2025. It lets non-programmers build working software quickly. Code made this way still needs review for errors and security problems.
Added October 21, 2026
Vision-language model — A vision-language model is a model that understands both images and text together. It can describe pictures, read screenshots, interpret charts, and answer questions about what it sees. These models power features such as photo search and visual assistants. They are a common type of multimodal model.
Added October 21, 2026
Voice cloning — Voice cloning is AI that copies a specific person's voice from a sample recording. The cloned voice can then say anything typed into the system. It has legitimate uses in accessibility, dubbing, and media. Criminals also use it in scams that imitate family members or executives.
Added October 21, 2026
W
Watermarking — Watermarking is embedding a hidden, detectable signal in AI-generated content to show where it came from. It can be applied to text, images, audio, and video. Watermarks help identify synthetic content and fight misinformation. Determined users can sometimes remove or weaken them.
Added October 21, 2026
Wayback Machine — The Wayback Machine is the Internet Archive's free service that stores old versions of web pages. It helps find old addresses, phone numbers, and names a business once published on its own website. Those old versions can explain where stale facts came from. It is a useful tool when hunting down drift.
Added October 21, 2026
WebPage (Schema.org) — WebPage is the Schema.org type for a single page on a website. It can state the page's name, description, date modified, and the main entity it is about. Connecting each page to the site's Organization ties content to identity. Many SEO plugins add WebPage markup automatically.
Added October 21, 2026
WebSite (Schema.org) — WebSite is the Schema.org type that describes a website as a whole, with its name and address. It is usually connected to the Organization that publishes it. It helps machines understand which business stands behind the site. Search engines also use it to show a site's name correctly.
Added October 21, 2026
Website Check — Website Check is a public tool that checks any website against public records. It shows the site's age, ownership, history, email setup, and warning signs of impersonation. It helps people judge whether a website is legitimate before trusting it. It is part of the audit step in the do-it-yourself path.
Added October 21, 2026
Weight — A weight is a parameter that sets how strongly one artificial neuron influences another. Training adjusts weights a little at a time until the model's answers improve. Together, a model's weights hold everything it has learned. Releasing a model's weights lets others run the model themselves.
Added October 21, 2026
WHOIS and RDAP — WHOIS and its modern replacement, RDAP, are services that show registration information for a domain, such as its registrar and creation date. Owner details are often hidden for privacy. Domain age and registration history help judge whether a website is established or newly created. Website Check uses these records to flag warning signs.
Added October 21, 2026
Wikidata — Wikidata is a free, open knowledge base of structured facts, run by the Wikimedia Foundation. Many search engines and AI/SI systems use it as a core reference for entities. Each item has a permanent identifier called a QID. Having an accurate Wikidata item, where a business qualifies, is one of the strongest identity anchors.
Added October 21, 2026
Wikidata QID — A Wikidata QID is the permanent identifier of an item in Wikidata, written as Q followed by a number. Linking a business's structured data to its QID through sameAs connects the business to the open knowledge graph. QIDs never change, even if names do. They are among the most valuable identifiers a business can link to.
Added October 21, 2026
Wikipedia — Wikipedia is the free online encyclopedia, and a major source of facts for knowledge graphs and AI/SI systems. Only notable organizations qualify for articles, under strict rules against self-promotion. Businesses should never write their own articles, but they can make sure verifiable facts exist elsewhere. Wikidata is often a more realistic identity anchor for smaller businesses.
Added October 21, 2026
WordPress — WordPress is the most widely used website platform in the world, powering a large share of all websites. Its plugin system lets owners add features without custom programming. PerfectReputations builds exclusively for WordPress websites. Every PerfectReputations premium tool installs as a plugin and works from the WordPress dashboard.
To learn more about WordPress, visit www.WordPress.org
Added October 21, 2026
Workflow automation — Workflow automation is using software, increasingly AI, to perform repeated business processes with little human effort. Examples include sorting email, processing invoices, and updating records. AI lets automation handle tasks that involve reading and judgment. It frees people for higher-value work.
Added October 21, 2026
X
X-Frame-Options — X-Frame-Options is a header that controls whether other websites may show a site's pages inside their own frames. Blocking framing prevents clickjacking, where a fake site overlays a real one to trick people. The SAMEORIGIN setting lets a site frame its own pages only. It is an easy, safe security improvement.
Added October 21, 2026
X-Robots-Tag — X-Robots-Tag is an HTTP header that carries robots instructions, such as noindex, from the server rather than from the page's HTML. It can apply to files like PDFs that cannot hold meta tags. Because it is invisible on the page, an accidental noindex here is easy to miss. Page Check checks this header on every page it scans.
Added October 21, 2026
xAI — xAI is an AI company founded by Elon Musk in 2023. It develops the Grok family of models, which are built into the X social media platform. The company has built very large computing clusters to train its models. It is one of several companies competing at the frontier of AI.
Added October 21, 2026
XML sitemap — An XML sitemap is a file that lists a website's pages so crawlers can find them all. It often includes the date each page last changed. Search engines use it to discover new and updated content quickly. Most SEO plugins create and update it automatically.
Added October 21, 2026
Y
Yelp — Yelp is a major review platform and a data source for other directories and services. Businesses claim their listing at biz.yelp.com to edit it. Its facts should match the canonical record, and duplicates should be reported. Yelp reviews also add independent corroboration of a business's existence.
Added October 21, 2026
Z
Zero-click answer — A zero-click answer is an answer a person gets directly from a search or AI/SI system without visiting any website. These answers are now common for questions about businesses. They make what the system says about a business more important than what the business's own pages say. Accurate identity data shapes zero-click answers.
Added October 21, 2026
Zero-click search — Zero-click search is a search in which the person gets an answer without clicking any result. AI answers and featured snippets make zero-click searches more common. This reduces visits to websites even when their content was used. It pushes businesses to care about how they are described, not only how they rank.
Added October 21, 2026
Zero-shot learning — Zero-shot learning is a model's ability to perform a task it was given no examples of, using only instructions. Asking a model to translate a sentence into French without showing any sample translations is zero-shot. Large models are surprisingly good at it. It shows how much general skill they gain from pretraining.
Added October 21, 2026
All definitions are copyrighted © 2026 PerfectReputations LLC.