A Theory of Canonical Identity

Identity, Not Discoverability: How AI/SI Systems Resolve, Trust, and Recommend Entities

By Prof. Emeritus Tim McGuinness, Ph.D. DFin.

October 19, 2026

Theory Abstract

The internet is moving from a discovery model built on search engines to one built on artificial intelligence and superintelligence systems, called AI/SI in this paper, that answer, synthesize, and recommend. Much of the search optimization industry has responded by renaming its existing services while continuing the same underlying work: keyword targeting, content production, and link acquisition. This paper proposes a theory that explains why that response aims at the wrong problem.

The Theory of Canonical Identity holds that AI/SI systems do not rank pages so much as resolve entities and then decide what they can confidently say about them. An entity that cannot be resolved to a single, coherent, corroborated identity cannot reliably be attributed facts, cited, or recommended, regardless of how much content it produces. Discoverability is therefore not a quantity to be optimized directly. It is an emergent property of identity integrity.

The paper defines the core terms of the theory, states its postulates and propositions, describes the mechanisms of synchronization, drift, conflation, and corroboration, and identifies six capabilities that current structured data vocabularies do not provide. It closes with implications for practice, testable predictions, and the theory’s limitations.

1. Introduction: From Retrieval to Resolution

Classical search was a problem of retrieval and ranking. A user typed a string, the engine matched it against an index of pages, and an algorithm ranked the candidates by relevance and authority, largely on the strength of keywords and the link graph. The optimization discipline that grew around this was rational for its medium. If visibility meant ranking a page for a query, the work was to produce pages that matched queries and to accumulate the signals that lifted them. The page was the unit of visibility, and the keyword was the unit of intent.

AI/SI systems work differently, and the difference is not cosmetic. When a user asks an AI/SI system who a company is, what it offers, whether it is reputable, or which provider to choose, the system does not return a ranked list for the user to evaluate. It composes an answer, acting as an oracle. To compose that answer, it must first resolve the entities the question refers to, recall or retrieve what it holds about them, judge how much of that it can trust, and then synthesize a response that attributes claims to the entities it has resolved.

The bottleneck is no longer retrieval against a string. It is the resolution of an entity and the confidence with which facts can be attached to it. This transition changes the unit of visibility from the page to the entity. The decisive question is whether the system can determine, without ambiguity, which entity a business is, distinguish it from every other entity it might be confused with, and attach to it facts held with enough confidence to state and recommend.

Entity resolution is not new. Statisticians formalized record linkage, the problem of deciding whether two records describe the same thing, more than half a century ago, and the study of entity resolution and knowledge graphs has grown into a substantial field. What is new is that this problem now stands between every business and its customers. The Theory of Canonical Identity applies the logic of entity resolution to the question of how businesses and other entities become visible, trusted, and recommended in the AI/SI era.

2. Definitions

The theory uses the following terms in specific senses.

Entity. Any distinct thing an AI/SI system can identify and describe, such as a business, an organization, a person, a product, or a place. This paper focuses on business and organizational entities, but the theory applies to any entity.

Identity assertion. Any statement, in any source, that attributes a fact to an entity: a name, an address, a relationship, a claim about what the entity is or does.

Representation. The model of an entity that an AI/SI system forms from all the identity assertions it has encountered, through training, retrieval, or both.

Resolution. The process by which an AI/SI system determines that a set of assertions refers to one particular entity, and not to another.

Canonical identity. A single, authoritative, internally consistent definition of what an entity is, expressed in a form both humans and machines can resolve, and designated as the source of truth from which every other representation of the entity derives.

Canonical record. The governed data record that holds the canonical identity: its identifiers, attributes, relationships, and the designation of its own authority.

Corroboration. The consistent assertion of the same identity by independent sources of differing authority.

Drift. The gradual divergence of an entity’s assertions across sources over time, as names, addresses, offerings, and people change and old assertions go uncorrected.

Conflation. The merging of two distinct entities into one representation, or the splitting of one entity into several.

Identity integrity. The degree to which an entity’s assertions, wherever they appear, are consistent with its canonical identity, corroborated by independent sources, and current.

Discoverability. The likelihood that an AI/SI system will correctly identify, accurately describe, and recommend an entity in response to a relevant request.

3. Postulates

The theory rests on five postulates about how AI/SI systems form and use their representations of entities. Each describes observable behavior of current systems and the established logic of entity resolution.

Postulate 1: Resolution precedes attribution. An AI/SI system must decide which entity a request refers to before it can attach facts to that entity. Facts cannot be reliably attributed to an entity the system cannot first resolve.

Postulate 2: Representation is aggregate. An AI/SI system forms its representation of an entity from everything it encounters about that entity across training and retrieval, not from any single privileged source. No one source, including the entity’s own website, controls the representation alone.

Postulate 3: Self-assertion is not authority. A claim an entity makes about itself carries no inherent authority, because anyone can assert anything about any entity. A claim gains weight only when independent sources corroborate it.

Postulate 4: Identity decays without maintenance. Entities change, and the assertions scattered across sources do not update themselves. Without active maintenance, an entity’s assertions drift apart over time.

Postulate 5: Confidence governs speech. An AI/SI system states and recommends what it holds with sufficient confidence, and hedges, omits, or misattributes what it does not. The confidence a system holds in an entity’s identity therefore limits what it will say about that entity.

4. Propositions of the Theory

From the postulates follow seven propositions. Together they form the Theory of Canonical Identity.

Proposition 1: Discoverability is an emergent property of identity integrity. An entity becomes discoverable when an AI/SI system can resolve it to a single coherent identity, corroborate that identity against independent sources, and attach facts to it with enough confidence to state and recommend. Each of these conditions is a property of the entity’s identity, not of any page. Improve identity integrity and discoverability follows. Pursue discoverability while the identity remains fragmented, and the effort treats a symptom while the cause goes untreated.

Proposition 2: Identity is upstream of content. Content acts on an entity’s representation only after the entity has been resolved. For a well-resolved entity, additional content adds facts the system can attribute with confidence. For a poorly resolved entity, additional content adds noise: every new claim from an entity the system cannot cleanly identify enlarges the surface over which the entity can be contradicted, conflated, or misattributed. Content amplifies the state of an identity. It does not repair it.

Proposition 3: Confidence follows consistency and corroboration, and contradiction outweighs agreement. An AI/SI system’s confidence in an identity rises with the consistency of its assertions and the independence and authority of the sources that corroborate them. A single contradiction can lower confidence more than many agreeing assertions raise it, because a contradiction forces the system to decide which version is true. An identity is only as coherent as its least consistent attribute.

Proposition 4: Boundaries are defended by distinction, not volume. The risk of conflation grows with an entity’s similarity to other entities, as with a common name, a shared founder or address, or a past rebrand, and shrinks with the strength of its distinguishing signals. The defense against conflation is the accumulation of distinguishing, corroborated attributes and relationships, not more content.

Proposition 5: Canonicality must be constructed, not declared. No current mechanism lets an entity simply declare one record the authoritative source to which all others must defer. Authority emerges from the pattern of stable identifiers, external linkage, and independent corroboration. Canonicality is therefore an engineered outcome, not a single property an entity can set.

Proposition 6: Identity integrity is a process, not a state. Because identity decays without maintenance, integrity achieved once does not persist. It is sustained only through continuous assertion, corroboration, and monitoring for drift and conflation, for as long as the entity exists. The earlier practice of keeping a business’s name, address, and phone number consistent across local directories was an early, narrow form of this principle. Canonical identity management is the full discipline that grew from that seed.

Proposition 7: Ambiguity is the attack surface for impersonation. Where an entity’s identity is fragmented, unverified, or weakly bounded, an impersonator’s assertions compete on nearly equal terms with the entity’s own, and an AI/SI system has no reliable basis for preferring the real one. A strong canonical identity, corroborated and verified at layers impostors cannot easily reach, such as authenticated email domains and independent registries, reduces the space in which impersonation succeeds. Identity integrity is therefore a security property as well as a visibility property.

5. The Structure of a Canonical Identity

A canonical identity has four components, and the theory treats each as a distinct source of consistency or contradiction. This paper identifies what a canonical identity requires. It does not attempt a step-by-step method, which depends on the entity, its history, and the platforms on which it appears.

One set of identifiers. The entity needs a stable canonical identifier that anchors it, linked to the authoritative identifiers that external knowledge systems already use. That linkage lets the entity’s assertion of itself be reconciled with the representations held in the major knowledge graphs. Without it, the entity floats free of the systems AI/SI models draw on to ground their answers.

One set of attributes. These are the facts that define the entity: its legal name, trading names, and aliases; its addresses and contacts; its founding and its people; its offerings and categories; and the claims it makes about itself. Each attribute is a potential point of consistency or contradiction, and under Proposition 3 the identity is only as coherent as its least consistent attribute.

One set of relationships. An entity is defined in part by its connections to other entities: its people, its parent and subsidiary organizations, its products, its memberships and affiliations, and the authoritative references that describe it. Relationships let a resolution system triangulate an entity against others and confirm that the entity being described is the one intended.

The designation of canonicality. The canonical identity includes the assertion that a particular record is authoritative and that all other representations defer to it. This component is the easiest to state as a principle and the hardest to achieve with current tools, for the reasons Proposition 5 and Section 7 describe.

6. Mechanisms: Synchronization, Drift, Conflation, and Corroboration

Under Postulate 2, a canonical identity that exists in only one place has not yet done its work, because AI/SI systems form their representations from everything they encounter. The identity must be asserted identically everywhere the entity appears, across the internet and throughout the enterprise itself. Four mechanisms govern whether it holds.

Synchronization is the assertion of the canonical identity, identically, in every place the entity appears: its own domains, its structured data, the major knowledge graphs, authoritative registries, verified profiles, and the independent references that describe it. The operative word is identically. Synchronization removes the contradictions that weaken a system’s confidence that scattered mentions refer to one entity. It is not a one-time markup exercise but an ongoing program, because of Postulate 4.

Drift is the failure synchronization prevents over time. As names change, addresses move, offerings evolve, and old profiles go unreconciled, an entity leaves a trail of stale and contradictory assertions that a resolution system must weigh against one another. Each uncorrected assertion is a standing contradiction, and under Proposition 3, contradictions weigh heavily.

Conflation is the failure synchronization prevents at the boundaries. When identity signals are too weak or too inconsistent to hold an entity’s boundaries firm, a system merges it with another entity or splits it into several. Businesses with common names, businesses that have rebranded, and businesses that share a founder or address with another are especially vulnerable. Under Proposition 4, the defense is rigorous, consistent identity assertion with strong distinguishing attributes, corroborated across independent sources, so that the entity’s boundaries are drawn the same way wherever a system looks.

Corroboration is the mechanism that turns consistency into trust. A single source asserting an identity is a claim. The same identity asserted consistently by independent sources of differing authority is a corroborated identity, and corroboration is what lets a resolution system move from recording a claim to trusting it enough to state and recommend. Corroboration is how the theory resolves Postulate 3: self-assertion becomes authority only when others confirm it.

7. Structured Data: Necessary and Not Sufficient

Structured data is the foundation of canonical identity, and Schema.org, expressed as JSON-LD, is its principal vocabulary. Schema provides the types that let an entity declare what kind of thing it is and the properties that let it assert its attributes and relationships in machine-readable form. Critically, it provides a node identifier that names an entity as a stable thing rather than describing it only in prose, and the property by which an entity links itself to its authoritative references elsewhere. That property is the primary tool available today for disambiguation and for reconciliation with external knowledge graphs. Any serious canonical identity program is built on a disciplined and complete implementation of this vocabulary.

Structured data is necessary, but it is not sufficient. Schema was designed to describe things, and it describes them well. It was not designed to govern identity, establish authority, or manage the trust, time, and provenance dimensions that identity demands in an adversarial, drifting, machine-resolved environment. The theory identifies six capabilities that the current vocabulary does not adequately provide.

First, authoritative canonicality. Schema lets an entity assert an identifier and link to its references, but it offers no strong, enforced way to designate one record as the canonical source to which all others defer, and no way to make that designation resolvable and trusted by consuming systems. Page-level canonicalization solved the analogous problem for duplicate web addresses years ago, but no entity-level equivalent carries the same force. This gap is the technical basis of Proposition 5.

Second, provenance and verification. Schema assertions are self-declared, and the vocabulary has weak facilities for expressing who asserts a fact, on what authority, and with what verification. It has no native trust layer. Trust must therefore be built externally, through corroboration across independent sources and through the verification mechanisms that particular platforms and registries offer, none of which the vocabulary coordinates.

Third, confidence and authority signaling. Resolution systems must weigh competing assertions, which requires knowing which are more authoritative. Schema offers almost no way to express the relative authority of sources or the ecosystem’s confidence in an assertion. This dimension exists only implicitly, in the pattern of corroboration across sources of differing authority, and it must be engineered deliberately.

Fourth, temporal validity. Identity facts change, and each has a period during which it is true. Schema’s handling of temporal validity is limited and inconsistent, so the history of an identity, its rebrands, relocations, and reorganizations, is poorly expressed. That history must be managed through disciplined versioning and reconciliation outside the vocabulary, or stale assertions will be treated as current.

Fifth, disambiguation by negation. The vocabulary offers a strong way to assert that two references are the same entity, but no equally strong, widely honored way to assert that two similar entities are distinct, which is precisely the assertion most needed to prevent conflation. The defense is therefore indirect, built from distinguishing attributes and corroborated linkages, as Proposition 4 describes.

Sixth, the retrieval and agent layer. The vocabulary describes an entity, but it does not govern how AI/SI systems retrieve, divide, weight, or attribute an entity’s content, or how an autonomous agent acting for a user should identify and trust the entity it transacts with. Emerging conventions that guide how language models read a site’s content now sit alongside the older conventions that governed web crawlers. This layer is new, unstandardized, and outside Schema’s scope, yet it is increasingly where resolution and attribution actually happen.

Taken together, these gaps mean that even a complete Schema implementation produces an entity that is well described but still self-asserted, still weakly protected against conflation, silent about the timing of its facts, unverified, and unlinked to any enforced notion of its own authority. Closing those gaps is the work of canonical identity management. Solutions exist for each of them, but they come from the deliberate engineering of corroboration, linkage, provenance, versioning, and monitoring, not from the vocabulary alone.

8. Implications for Practice

8.1 A Discipline, Not a Project

If the theory holds, canonical identity management is a continuous, engineered program rather than a one-time campaign. Its elements follow directly from the propositions.

Establish the canonical record. Define the single authoritative and internally consistent identity of the entity, with its identifiers, attributes, relationships, and the designation of which record is the source of truth. This is a data modeling exercise, and its output is a governed record, not a web page.

Express it in structured data. Implement the identity completely and correctly in the vocabulary, with stable identifiers and disciplined linkage to the authoritative external references that let the entity be reconciled with the major knowledge graphs.

Corroborate it across independent authorities. Assert the same identity consistently across the entity’s own domains, the knowledge graphs, authoritative registries, and verified profiles, so that the identity is confirmed from many directions rather than merely claimed.

Verify it where impostors cannot reach. Anchor the identity in layers that require control of the entity’s own infrastructure, such as authenticated email domains with enforced policies, so that the identity resists impersonation as well as confusion.

Close the vocabulary’s gaps deliberately. Engineer the provenance, verification, temporal validity, disambiguation, and authority dimensions that structured data does not express, through linkage, corroboration, versioning, and platform verification.

Develop what the stack still lacks. The full set of layers needed to model, assert, verify, reconcile, and maintain a canonical identity does not yet exist in complete or standardized form. There is no enforced mechanism for canonical authority, no shared trust layer across platforms, no consistent expression of provenance and temporal validity, and no settled standard for how autonomous agents identify and trust entities. Part of the discipline is therefore research and development: prototyping the identifiers, methods, and protocols that close these gaps and pressing them toward adoption. In a transition moving this quickly, waiting for the market to supply them is itself a decision to remain unresolved.

Govern the retrieval and agent layer. Manage the emerging conventions that guide how AI/SI systems and autonomous agents read, trust, and attribute the entity’s content, and decide which systems may use that content, and for what purposes.

Monitor and reconcile continuously. Treat the identity as a living asset that degrades without care. Watch for drift, contradiction, stale assertions, conflation, and impersonation, and reconcile them as they appear.

8.2 Why Relabeled Search Optimization Misreads the Problem

Much of the search optimization industry has met this transition by renaming its services, calling the same work generative engine optimization, answer engine optimization, AI visibility, or AI discoverability, often with structured data added as a minor ingredient. Underneath the new labels, the work remains keyword targeting, content production, and link acquisition.

The theory explains why this approach fails. It applies a page-and-keyword model to a problem that concerns neither pages nor keywords. Under Proposition 2, content optimization operates downstream of entity resolution: when an AI/SI system cannot confidently resolve who a business is, more content does not help and can actively hurt, because each new claim from a poorly resolved entity widens the surface for contradiction and conflation. Producing content for an unresolved identity is like pouring water into an unsealed vessel.

The deeper issue is one of discipline and skill. The work the theory describes is data modeling and knowledge engineering, drawing on structured data, entity reconciliation, graph thinking, and provenance. None of these are the native competencies of a content marketing or link-building firm. When such a firm renames itself an AI visibility agency, it changes its signage without acquiring the knowledge the new signage implies. The mechanics it does not understand are not an advanced version of the mechanics it knows. They belong to a different field and a different era.

9. Testable Predictions

A theory earns its standing by making predictions that observation can confirm or refute. The Theory of Canonical Identity makes the following predictions, each of which can be tested by querying AI/SI systems about matched sets of entities and measuring how accurately and how often those systems identify, describe, and recommend them.

Prediction 1. Among comparable entities with similar content volume, those with consistent, corroborated identities will be described more accurately and recommended more often by AI/SI systems than those with fragmented identities. This follows from Propositions 1 and 3.

Prediction 2. For entities with poorly resolved identities, increasing content volume without improving identity integrity will not improve, and may reduce, the accuracy with which AI/SI systems describe them. This follows from Proposition 2.

Prediction 3. Entities with common names, recent rebrands, or shared founders or addresses will show higher rates of misattribution and conflation in AI/SI answers, and those rates will fall measurably after their identities are reconciled and their distinguishing attributes corroborated. This follows from Proposition 4.

Prediction 4. Correcting drift, by removing or reconciling stale and contradictory assertions, will produce a measurable improvement in the accuracy of AI/SI descriptions, and the improvement will exceed what an equal number of new, consistent assertions would produce. This follows from the asymmetry in Proposition 3.

Prediction 5. Without ongoing maintenance, the accuracy of AI/SI descriptions of an entity will decline over time as the entity changes and its assertions drift, even when no new contradictory information is deliberately introduced. This follows from Proposition 6.

Prediction 6. Entities with verified, corroborated identities, including authenticated email domains with enforced policies, will experience fewer successful impersonations, and AI/SI systems will be less likely to surface impostors in their place. This follows from Proposition 7.

10. Limitations and Open Questions

The theory has limits that future work must address. First, AI/SI systems are largely opaque. Their training data, retrieval methods, and confidence thresholds are not published, so the theory’s postulates describe observed behavior and the established logic of entity resolution rather than the documented internals of any particular system. Different systems may weigh sources differently, and the same system may change its behavior without notice.

Second, measurement is difficult. Testing the predictions requires repeated, controlled queries across systems and over time, with matched entities and agreed standards for what counts as an accurate description or a recommendation. Those methods are still being developed.

Third, the standards are moving. Structured data vocabularies, conventions for guiding language models, and protocols for autonomous agents are all evolving. Some of the gaps described in Section 7 may narrow as new standards emerge, and new gaps may open. The theory’s central claim, that identity integrity governs discoverability, should hold even as the tools change, but its specific prescriptions will need revision.

Finally, the theory addresses entities that want to be found and trusted. It says less about entities that deliberately obscure their identities, about the ethics of AI/SI systems that decide which entities to recommend, or about how individuals, as distinct from organizations, should manage their own identities. These remain open questions.

11. Conclusion

Businesses now paying relabeled search agencies for AI visibility are, in most cases, buying content optimization for a problem that is not a content problem. The Theory of Canonical Identity holds that whether an entity is found, understood, and recommended by AI/SI systems depends on whether those systems can resolve it to a single, coherent, corroborated, trusted identity. That is an identity problem. Its solution is the establishment and rigorous synchronization of a canonical identity, founded on structured data and extended past the limits of the current vocabulary into provenance, verification, temporal validity, disambiguation, and authority.

Discoverability, in the end, is not the thing to pursue. It is the reward for doing the identity work correctly. The practitioners who understand this are not doing a more advanced form of search optimization. They are doing a different kind of work entirely: the engineering and governance of machine-resolvable identity. In the transition now underway, that work will separate the entities AI/SI systems can confidently speak about from those they cannot resolve, cannot trust, and therefore cannot recommend.

Copyright © 2026 Tim McGuinness, Ph.D. All rights reserved.

References