A brand earns coverage in fifty articles over two years. Good sources, reputable outlets, strong editorial context. By any reasonable measure, its AI visibility should be building.
But when you test the brand name in ChatGPT, Gemini, and Perplexity, it barely appears. When it does appear, the answers are inconsistent — right in one query, wrong in another, absent in a third.
The coverage is real. The problem is that the AI cannot tell it is all about the same brand.
This is identity drift — what Kurt Fischman of Growth Marshal calls the “silent killer” of AI brand visibility. Silent because no individual piece of content is wrong. Killer because the cumulative effect is invisible until you run the numbers. And it almost certainly affects your brand right now.
What Identity Drift Is
Identity drift occurs when a brand appears under multiple name variants across its published content and structured data, without explicit canonical links connecting those variants to a single entity.
An example. A brand’s published content uses:
- “Entity Confidence” (formal brand name)
- “EC” (abbreviation, used informally in team posts)
- “entityconfidence.ai” (domain name, used as identifier in some PR coverage)
- “the EC team” (used in founder-authored content and podcast appearances)
From a human reader’s perspective, these are obviously all the same thing. From an AI retrieval system’s perspective, each unlinked variant is a candidate for a separate entity. Evidence accumulates under four separate nodes. None achieves the recognition threshold that the consolidated entity would have reached.
The error is structural, not factual. No individual page contains wrong information. The damage is done by the aggregation — fifty articles of genuine coverage, split four ways, producing four weak signals instead of one strong one.
Why the AI Cannot Just Figure It Out
Understanding identity drift requires understanding how AI systems resolve entity mentions — and why the process breaks down when naming is inconsistent.
AI retrieval systems run a five-stage entity resolution pipeline:
Stage 1 — Named Entity Recognition. Every text span is classified: is “EC” an Organisation, a Location, an abbreviation? Consistent capitalisation, explicit schema @type declarations, and contextual identifiers like “Ltd” or “Inc.” all raise the confidence of this classification.
Stage 2 — Entity Linking. The recognised mention is mapped to a specific canonical entry in the model’s knowledge base. This is where sameAs properties in your JSON-LD do their work — they explicitly tell the model that your website entity is the same as the Wikidata entry, the same as the LinkedIn Company Page, the same as the Crunchbase profile. Without these links, the model is working from text alone.
Stage 3 — Attribute Verification. Once linked, the model retrieves attributes: products, leadership, locations, capabilities. Confidence in each attribute depends on source agreement across the web. When your website, Wikidata entry, and LinkedIn page all state the same founding date and CEO name, confidence is high. Conflicting sources reduce it.
Stage 4 — Temporal Validation. Freshness is evaluated. Stale data that conflicts with more recent signals reduces overall confidence, regardless of original accuracy.
Stage 5 — Cross-Reference Verification. Claims are compared across multiple independent sources before recommendation. Entities with consistent, corroborated data across independent sources achieve the highest confidence scores. The output is a probability distribution, not a binary pass/fail.
Identity drift attacks Stage 2. When a brand appears as “Entity Confidence,” “EC,” “entityconfidence.ai,” and “the EC team” without sameAs links connecting them, the entity linking step fails to consolidate evidence. For a brand with a common or abbreviated name, “EC” could be East Coast, Ecuador, or any of a dozen other organisations. Without structured data linking “EC” to the same canonical identifier as “Entity Confidence,” the model has no reliable basis for consolidation.
The system is not being careless. It is designed to prioritise confident entity resolution. A brand with fragmented naming is, by design, deprioritised in favour of a brand that has made its identity easy to resolve.
The Compounding Problem
Identity drift worsens over time. A small team publishing consistently will accumulate drift across a growing content catalogue — blog posts use the abbreviation, press releases use the formal name, podcast appearances use whichever comes naturally in context.
No individual page creates a visible problem. But the ratio of canonical to off-canonical mentions shifts steadily. At some threshold, the evidence fragmentation is severe enough to suppress recommendation frequency across all variants.
The three states:
Single canonical name, consistent sameAs links: evidence consolidates under one entity node; salience rises; AI recognition is reliable.
Multiple variants without canonical linking: evidence fragments; each variant falls below the recognition threshold; AI confidence is low.
Mixed (canonical plus some drift): partial consolidation; recommendation frequency varies by query phrasing; visibility is inconsistent and difficult to diagnose.
Most brands with an identity drift problem are in the mixed state. They can be recommended — and sometimes are — but the inconsistency makes it look like a platform-specific or query-specific problem when it is actually a naming problem.
Conflation: Worse Than Invisibility
Incomplete disambiguation produces a second failure mode beyond low recommendation frequency: conflation.
Conflation is when the AI resolves a name to the wrong entity — a different company with a similar name, a subsidiary with an overlapping product line, an acquired brand that no longer exists independently. The model is not refusing to answer. It is answering confidently with information about a different brand.
This is harder to detect than absence. A brand that is absent from AI answers generates a clear gap. A brand that is conflated with a competitor generates confident, plausible-sounding wrong answers — wrong founding date, wrong product description, wrong market position — that most users and many brand teams will not notice unless they are specifically auditing AI output.
The common triggers: generic or widely shared brand names; trading name used publicly alongside a different legal entity name; acquired companies whose independent web presence still ranks well; regional subsidiaries operating under slightly different identifiers. Each inconsistency raises the probability that the model will hedge or conflate rather than identify it cleanly.
The Audit
Four steps to establish whether identity drift is suppressing your brand’s AI visibility:
Step 1 — List all name variants in active use. Audit published pages, structured data, social profiles, PR coverage, and team-authored content for every form in which the brand is referenced. Include abbreviations, domain name references, legacy names, and product variants. The list is typically longer than expected. Any variant appearing in third-party coverage without a canonical sameAs link connecting it to your registered entity is a fragmentation risk.
Step 2 — Check sameAs coverage in JSON-LD. Every page that includes Organisation schema should have sameAs links pointing to the same canonical external identifiers: Wikidata entity URL, LinkedIn Company Page, Wikipedia article (if applicable), Companies House or equivalent, Crunchbase profile. Inconsistency across pages introduces its own disambiguation noise. Missing sameAs links leave the model without structured evidence to resolve variants.
Step 3 — Run Google NLP API entity extraction on key pages. The API returns exact salience scores and entity classifications for the content it analyses and uses the same NLP infrastructure as Google Search and AI Overviews. A well-optimised page should show 2–3 primary entities with scores above 0.5, supporting entities clustered at 0.2–0.4, and unrelated entities absent. If the brand’s canonical name is returning a salience score below 0.5 on its own About page, that is a measurable identity problem. If abbreviations or domain variants are returning as separate entities, drift is confirmed.
Step 4 — Test canonical and variant names in major AI engines. Query ChatGPT, Gemini, and Perplexity with both the canonical name and common variants. Compare the responses: is representation consistent? Does one variant produce different — or absent — answers compared to another? Inconsistency between query variants points to fragmented entity evidence.
The Fix
Establish one canonical name and use it consistently. Choose the form that appears most clearly in official registration, structured data, and high-authority third-party sources. Define it formally — a one-line canonical entity statement — and use it as the reference for all new content.
Add sameAs links to all entity declarations. Every page with Organisation schema should link to the same canonical external identifiers. Person schema for founders and key executives should link to LinkedIn profiles with worksFor connecting back to the Organisation. The model uses these relational signals to build a structured graph of who you are and who is associated with you.
Audit and prioritise existing content. Full remediation of every off-canonical reference in an existing content archive is rarely practical immediately. Prioritise the highest-traffic pages, the most-linked-to pages (check backlink data), and pages that rank for queries where you want AI recommendation. Update their canonical name usage and structured data first.
Build an entity registry. A one-page internal document per entity — canonical name, description, Schema.org type, official identifiers (Wikidata QID, Wikipedia URL, LinkedIn URL, company registration number) — that gates all new content publication. Before any new page is published, the entity references on it should be checked against the registry. This prevents drift from accumulating in the first place.
The evidence that this works is concrete. Schema App found that scaling entity linking through consistent sameAs properties and NAP data produced a 46% increase in impressions and a 42% increase in clicks for non-branded queries — direct evidence that disambiguation signals have measurable reach effects, not just AI-recommendation effects.
Building Salience Once Drift Is Fixed
Eliminating identity drift is the prerequisite. Once evidence is consolidating under a single canonical entity, the next step is building the salience of that entity in content.
Entity salience — the NLP score (0–1) measuring how central a named entity is to a document — is not the same as mention frequency. Repeating a brand name ten times without a clear semantic frame produces lower salience than using it once in the subject position of a strong opening sentence.
Growth Marshal’s Kurt Fischman describes a four-stage operational sequence:
Stage 1 — Define canonical entities. Create a one-page brief per entity: canonical name, description, Schema.org type, key properties, and official identifiers (Wikidata QID, Wikipedia URL, LinkedIn URL). Keep URLs stable. Gate all new content publication on this registry.
Stage 2 — Rewrite core pages. Lead with identity, explain the role, show relationships. Keep paragraphs tight. Remove metaphors that introduce ambiguity. Each paragraph should function as a semantic capsule: a subject-verb-object sentence that pins the entity to an attribute, maintained proximity between entity and defining content, a restatement at the close. Every sentence should advance the model’s understanding of which entity matters and why.
Stage 3 — Mirror the narrative in JSON-LD. Use Organization and Person for the firm and leadership; Product or Service for offerings; CreativeWork for papers and guides; DefinedTerm for proprietary concepts. Link sameAs to official profiles. The markup should translate what the prose already says — not invent a different story.
Stage 4 — Measure and iterate. Run entity extraction monthly via the Google NLP API. Track salience ranks and spreads: the primary entity should rank first by a wide margin; secondary entities should cluster below; unrelated entities should not appear at all. Test prompts in major AI assistants to observe recommendation lift.
The signals with the highest individual impact: subject position (“Entity Confidence is…” assigns maximum salience; “the concept introduced by…” does not); H1 prominence (the heading is the strongest document-level salience signal); and first-100-word placement (entities introduced in the opening paragraph receive higher salience than those introduced later).
The Sequencing Rule
Lumar’s GEO framework makes a structural point that applies directly here: entity GEO must precede brand authority GEO.
Before an AI platform can decide whether to trust or recommend a brand, it must first know who the brand is — which specific organisation, product, or person the content refers to. Brand authority signals, earned media coverage, E-E-A-T indicators — these have nothing to attach to if entity recognition has not been established. They accumulate under weak, fragmented nodes and fail to translate into recommendation.
The ordering matters because it determines which work generates compounding returns. Identity clarity is the prerequisite layer. Built correctly, it makes everything else — earned coverage, structured content, cross-platform consistency — contribute to the same entity node. Built incorrectly, or left unbuilt, it causes everything else to work against a fragmented target.
This is why identity drift is the first audit, not the last. Not because it is the most sophisticated problem, but because every other investment in AI visibility is contingent on solving it first. The brand that fixes its entity layer before adding more content will compound every piece of coverage it has already earned. The brand that skips this step will keep publishing into a leaking bucket.
Fix the container before you fill it.