A company can have a technically sound website, accurate product pages and a clean content audit while an AI assistant quietly describes it using facts that belong somewhere else. When the model has too little reliable information about the company itself, it may complete the picture with a better-documented competitor, a category norm or an older version of the brand.
That is the central argument in Duane Forrester's September 3 Search Engine Journal analysis. The problem is important because it changes where AI visibility teams need to look for errors. A conventional content audit inspects what a company has published. The substitution Forrester describes becomes visible only in the model's output — often in response to questions the company never sees.
The idea should be handled with one important qualification. Forrester is presenting an operational framework informed by research on long-tail factual knowledge, model memory and entity retrieval, not a new controlled study showing that a specific percentage of sparse brands are replaced by competitors. The underlying weaknesses are documented; the commercial-brand substitution pattern is a practitioner interpretation of how those weaknesses can surface in real AI answers.
Even with that distinction, the problem is highly relevant to AI visibility. Being discoverable is not enough if the system reconstructs the wrong entity once it finds you.
Sparse entities are harder for language models to represent
Large language models do not possess equally detailed knowledge about every company, person or product. Popular entities appear repeatedly across training data, public databases, journalism, reviews, documentation and the broader web. Smaller or less documented entities occupy a thinner part of that information distribution.
Forrester points to research showing that factual recall deteriorates for less popular knowledge and that simply increasing model scale does not eliminate the long-tail problem. Bigger models become better at recalling common facts, but obscure entities remain comparatively difficult.
That matters for a regional services firm, a mid-market software vendor or a young publisher. The model may know the category extremely well while knowing very little about the specific company being discussed.
When an answer still has to be produced, the system can lean on stronger patterns surrounding the weak entity. The result may sound coherent because the borrowed information is plausible for a company of that type.
The dangerous error is often plausible rather than absurd
AI hallucinations are often illustrated with obviously fabricated citations, invented statistics or nonexistent people. Brand substitution can be harder to detect because the answer may be reasonable.
A project-management company might be described as offering a standard pricing model used by its largest competitor. A cybersecurity vendor might inherit an implementation timeline common to the category. A software company that changed positioning two years ago may still be described using its previous product architecture.
None of those answers needs to look ridiculous. In fact, plausibility is what makes the failure commercially dangerous. A buyer unfamiliar with the company has no reason to know that the description belongs to a neighboring entity or an outdated version of the brand.
Forrester separates this from purely random fabrication and describes the behavior as directional: when evidence about one entity is sparse and evidence about a nearby entity is dense, the answer can drift toward the stronger information pattern. That is a useful diagnostic model, although the degree to which this occurs consistently across brands and models still requires broader empirical measurement.
Four different failures can look like the same confident answer
Forrester identifies four practical forms of substitution. The first is silent analogy, where the model fills missing company-specific information with characteristics of a well-documented competitor or a typical company in the same category. Pricing, features or implementation details can be borrowed without any visible indication that an inference occurred.
The second is staleness. The model has genuine information about the company, but the information describes an earlier version of it. A discontinued service, departed executive or abandoned positioning statement can be presented in the present tense because the model does not automatically attach an expiration date to every fact stored in its parameters.
The third is confidence built on thin evidence. A company supported by one weak external source can be described in the same authoritative tone as a company supported by dozens of independent references. The generated prose usually does not expose how much evidence sits underneath each statement.
The fourth is category substitution. The model understands the industry and answers from that knowledge, then attaches the resulting generalization to the specific company. This may be the hardest version to catch because the statement can be true about the market while still being unsupported or false for the brand.
A clean website audit cannot detect an answer generated elsewhere
This is where the problem breaks with traditional SEO workflows. A content audit can identify missing pages, stale copy, contradictory product descriptions, broken schema or poor internal linking. Those are all useful checks, but they inspect a surface the company owns.
An AI answer is assembled on another platform from some combination of model memory, retrieved documents, search results and system-specific ranking or synthesis processes. A company can inspect every sentence on its website and still never encounter the inaccurate description a prospective customer receives in ChatGPT, Gemini, Claude or another assistant.
The absence itself is the problem. The evidence needed to prevent a wrong answer may not exist in enough places, may not be retrievable for the relevant query or may be overwhelmed by older and more widely repeated information.
This is closely related to a distinction NetContentSEO has been testing with small publishers. In our experiment on whether an LLM can recognize a small publisher before Google fully trusts it, recognition and accurate reconstruction are deliberately treated as separate events. A system can identify the correct entity while still misunderstanding what that entity represents.
Publishing more is useful, but it is not automatically a cure
The intuitive response to weak AI knowledge is to create more content. In many cases that is sensible. Clear product pages, current documentation, explicit company descriptions and consistent entity information give retrieval systems better material to work with.
Forrester argues, however, that retrieval itself can reproduce popularity bias. He cites research on entity-centric questions showing that dense retrieval systems can perform worse on uncommon entities and generalize more reliably to popular ones. In other words, a model with access to retrieval does not automatically escape the same long-tail problem that affects what it learned during training.
This does not mean retrieval is useless or that publishing authoritative information cannot help. It means “write the missing page and the model will retrieve it” is too deterministic. The page must be discoverable, understood as belonging to the correct entity, considered relevant to the specific question and selected over competing evidence.
A weak brand representation is therefore not always a page-count problem. It can be an evidence-distribution problem.
Third-party descriptions can become part of the brand's machine identity
Companies naturally focus on the content they control, but AI systems can form their picture of an entity from information distributed across the web. Industry publications, directories, partner pages, reviews, interviews, documentation repositories and comparison pages may all contribute to the available evidence.
If those external descriptions are sparse, contradictory or old, a brand can have an accurate website while still presenting an unstable machine-readable identity.
This creates an uncomfortable implication for communications and digital PR. External coverage is not valuable only because it produces links or human awareness. It can also help establish independent descriptions of what the company is, what it offers and how it differs from neighboring entities.
That does not justify manufacturing mentions or flooding the web with repetitive brand claims. Artificial repetition can create its own quality problems. The objective is consistent, verifiable information appearing in contexts that make sense for the company.
The outdated version of your brand may have more evidence than the current one
Staleness deserves particular attention because companies frequently change faster than their external information footprint. A rebrand can happen in weeks. A product can be discontinued overnight. Pricing and positioning can change in a quarter. The old version may remain documented across years of articles, directories, archived pages and model training data.
From a machine perspective, the historical description can therefore have more supporting evidence than the new one. Updating the homepage does not erase the previous entity representation from the rest of the web or from a model's parametric memory.
Teams should treat major company changes as information migrations, not merely website edits. New positioning needs enough consistent evidence to replace the old pattern across surfaces that matter.
This is one reason AI answers can expose organizational history that marketers thought they had already retired.
Category averages are especially difficult to notice
A competitor's feature incorrectly attributed to your company can eventually be spotted. Category knowledge is more subtle. If most enterprise software in a market takes several weeks to implement, an AI system may describe that timeline as though it applies specifically to one vendor with little public implementation data.
The answer sounds credible because the statement is plausible for the category. A casual fact check may not flag it. Yet the company may have a radically different onboarding model.
The same pattern can affect pricing structures, contract terms, integrations, customer profiles, geographic coverage and service models. Whenever a brand differs materially from the category norm, the difference needs unusually clear evidence because the model has a strong generic pattern available as a fallback.
Distinctiveness therefore creates an information obligation. The more a company deviates from what is typical in its market, the more explicitly that deviation needs to be documented.
AI visibility audits need to inspect outputs, not only inputs
Forrester's most actionable recommendation is methodological: test the generated answers themselves. A traditional audit asks whether the website contains accurate information. An AI visibility audit should also ask what models say when the company is not controlling the context.
The obvious prompt — “Tell me about Brand X” — is useful but insufficient. Naming the company gives the retrieval system a strong entity cue. More revealing tests can use category questions, capability questions, comparison prompts or prompts naming only a competitor and asking for alternatives.
Those conditions force the system to decide whether the brand belongs in the answer and which facts it associates with the brand once selected. That is where weak entity representations and borrowed descriptions are more likely to become visible.
The tests also need repetition. Generative outputs vary between runs, and retrieval indexes and models change over time. As NetContentSEO has argued in our broader work on being understood, validated and cited by AI, recognition alone is not a sufficient success metric. Correct reconstruction needs to be observed repeatedly.
Trace unsupported claims back to the evidence gap
The useful unit of analysis is not simply whether an answer is “good” or “bad.” Teams should isolate specific claims and ask where each one came from.
If a model says the company offers a feature it does not have, the immediate task is to identify whether the statement resembles a competitor, an old product version or a category convention. If it quotes an outdated employee count, find the sources still reinforcing that number. If it consistently misstates the target customer, compare the company's current messaging with external descriptions that remain indexed.
This turns an AI error into an evidence map. The remediation might involve clearer first-party documentation, updated directory profiles, corrections to third-party pages, better entity consistency or new authoritative coverage.
The goal is not to manipulate the model directly. It is to reduce the amount of ambiguity the model encounters when it tries to reconstruct the company.
One wrong answer is not enough to diagnose a systematic problem
Generative models are probabilistic. They can produce an inaccurate answer once and a correct answer on the next run. That makes anecdotal screenshots useful for discovering failure modes but weak as evidence of their frequency.
A serious monitoring program should use a stable prompt set, repeat tests, record model versions and retrieval conditions where possible, and distinguish isolated hallucinations from recurring substitutions. Cross-model comparison is also useful because the same brand may be represented well in one system and poorly in another.
This is especially important because Forrester works commercially in AI visibility measurement and explicitly discloses that interest in his article. The framework should therefore be evaluated on its evidence and reproducibility rather than treated as a neutral industry benchmark simply because it identifies a plausible risk.
The strongest case for action is a repeated pattern: multiple runs or systems producing the same unsupported claim, especially when that claim can be traced to a stronger neighboring source.
AI visibility is partly a battle against ambiguity
Traditional SEO often asks whether a page can be crawled, indexed and ranked. AI visibility adds another layer: whether a system can form a stable and accurate representation of the entity behind those pages.
For established brands, that representation may be reinforced by thousands of independent sources. For smaller companies, the evidence can be thin enough that the model has to interpolate. That is where competitors, category averages and stale descriptions become dangerous substitutes.
The practical response is not to publish indiscriminately. It is to make important facts explicit, current and consistent; strengthen independent evidence where it genuinely belongs; and monitor model outputs in the kinds of questions buyers actually ask.
A company does not need an AI system to know everything about it. It needs the system to know enough that, when evidence runs thin, the nearest competitor does not become the easiest answer.