Does AI actually know what your company does?

Ask an assistant what your company does and you may not like the answer. It might describe you as you were two years ago, or as a different kind of business than you are, or in terms so generic they would fit any competitor. Marketing leaders often read this as a content problem and respond by publishing more. It is usually not a content problem. It is an entity problem, and publishing more content on a confused identity tends to deepen the confusion rather than resolve it.

If an AI assistant describes your company vaguely, out of date, or as the wrong kind of business, the issue is entity understanding: the model has not formed a clear, confident picture of who you are. It assembles that picture from how consistently you are described across your own site, your structured data, and the third-party sources that mention you. When those disagree, you present as a fuzzy entity, and a model reaches for a clearer competitor over a blurry you, because it will not stake a recommendation on an identity it cannot pin down. This article explains how that picture gets built, how to tell entity confusion apart from other visibility problems, and how to fix it.

What an entity is, and why it decides your visibility

An entity, in this context, is the thing a model believes your company is: a distinct, identifiable organisation with a name, a category, a set of attributes, and relationships to other entities. It is not a page and not a keyword. It is the model's internal answer to "who is this".

This matters because AI systems increasingly reason about the world in terms of entities rather than strings of text. When a buyer asks for "a consultant who helps hotels stay visible in AI search", the assistant is not matching that phrase against your homepage word for word. It is trying to identify entities that fit the description, a person or company whose known attributes include the category, the specialism, and the audience. If your entity is clear and its attributes line up with the request, you are a candidate. If your entity is fuzzy, or its attributes are contradictory across the sources the model has read, you are passed over in favour of a company the model is more confident about. Confidence, not just relevance, is what earns the recommendation.

That is the uncomfortable part. A model would rather name a slightly less perfect fit it understands clearly than a perfect fit it is unsure about, for the same reason a person recommending a supplier reaches for the one they can describe confidently. Ambiguity is not neutral. It actively costs you the recommendation.

The symptoms of a weak or confused entity

Entity problems announce themselves in specific ways once you know what to look for, and they are distinct from the symptoms of a retrieval or extraction problem.

The clearest symptom is being described inaccurately when you are named. The model knows you exist but gets you wrong: it names an old product, an outdated positioning, a service you no longer offer, or the wrong category entirely. That is not a page it failed to read. It is a picture it assembled from stale or conflicting sources. A second symptom is vagueness: when asked about you directly, the assistant produces something generic that could describe any firm in your space, which means it has your name but few confident attributes attached to it. A third is inconsistency across surfaces: ChatGPT describes you one way, Gemini another, Google's Overview a third, because each has read a different slice of a web that does not agree about you. And a fourth, subtler one, is being absent from category questions you clearly fit while still being findable by name, which tells you the model can retrieve you when handed your name but has not associated you with the category you belong to.

Contrast these with a retrieval problem, where you are simply absent everywhere, name included, because your pages are not being reached at all. Telling the two apart matters, because the fixes are completely different, and the manual check that separates them is set out in how to check whether your brand appears in ChatGPT and AI answers. Entity problems are the ones where you appear but wrongly. Retrieval problems are the ones where you do not appear at all.

Where the picture comes from

To fix how a model understands you, you have to know what it reads, and the answer is: far more than your own website.

A model's picture of your entity is built from the accumulation of how you are described everywhere. Your own site and its structured data are part of it, and the part you fully control, but they are not the whole. Your social profiles, your listings in directories and industry databases, the way journalists and bloggers refer to you, the descriptions in roundups and comparisons, the way people talk about you in communities and forums, all of it feeds the picture. When these sources agree, the entity is sharp and the attributes are confident. When they contradict each other, the model is left to reconcile competing claims, and the safest thing it can do with an unreconcilable identity is decline to stake much on it.

This is why the instinct to fix entity problems by rewriting your homepage only goes so far. Your homepage is one voice in a chorus. If the rest of the chorus is singing an old positioning, a clearer competitor, or nothing at all, updating your own voice helps but does not settle the question on its own. The work is to make the whole chorus consistent, starting with the parts you own and extending to the parts you can influence.

How to make your entity unmistakable

The fix is consistency, applied deliberately across the sources a model reads, in roughly the order of how much control you have.

Start with the assets you own outright. Make your name, your category, and your one-line description identical across your homepage, your about page, your footer, and your metadata. Pick one way of saying what you are and use it everywhere, rather than a different clever phrasing on each page. State the category in plain words, not only in a tagline: if you are a consultant, the word "consultant" should appear where it matters, because a model cannot infer a category you never name. Remove the internal contradictions first, because a site that describes itself three ways is teaching the model to be unsure about you before any third party gets involved.

Then add the machine-readable layer. Implement correct Organization and Person structured data, with sameAs links pointing to your real, consistent profiles, so the model has an explicit aid to connecting your various pages and profiles to a single identity. This is genuinely useful, but it is a comprehension aid, not a magic assertion, and it will not override a web full of contradictions on its own. Why schema helps less than people expect, and what it can and cannot do, is worth understanding in its own right, which is the subject of why your schema markup isn't helping AI understand you.

Then extend consistency outward. Update the descriptions in the directories, profiles, and databases that describe you, so the third-party picture matches the one you own. This is slower and only partly in your control, but it is where a lot of stale entity information lives, and where a model often gets its outdated impression. The goal throughout is singular: everywhere your company is described, it should be described the same way, so the model assembling your entity finds agreement instead of a puzzle.

Give the model attributes worth attaching

Consistency removes confusion, but a clear entity with thin attributes is still a weak candidate. The second half of entity work is giving the model specific, distinctive things to know about you.

A model recommends confidently when it can attach concrete attributes to your entity: the exact category you serve, the audience you specialise in, the method you use, the results you are known for. Vague positioning gives it nothing distinctive to hold, so you blur into the category average. Specific positioning, stated plainly and corroborated by others, gives the model reasons to surface you for the precise questions you are the best answer to. This is where entity work meets originality: the companies that are described confidently and specifically are usually the ones that have given the world something specific to say about them, a named approach, original data, a clear specialism, rather than a generic claim to be good at everything.

The corroboration of those attributes by independent sources is what makes them stick, because a model weighs what others say about you more heavily than what you say about yourself. Being described as a specialist in your niche by a source the model trusts does more for your entity than any amount of self-description, which is the same reason a competitor with genuine third-party standing keeps getting recommended ahead of you, covered in why AI recommends your competitor instead of you.

Why this gets worse after a rebrand or repositioning

Entity problems are at their most acute in the year or two after a company changes its name, its category, or its positioning, and understanding why explains a lot of otherwise puzzling AI descriptions.

When you reposition, you update the sources you control quickly: your homepage, your profiles, your metadata. But the wider web updates on its own schedule, or not at all. The old directory listing, the years of articles describing your previous positioning, the cached descriptions and third-party mentions all still carry the former identity. A model reading across that mixture sees a company described one way by its own current site and another way by the accumulated record, and the accumulated record is larger. So it hedges, or it defaults to the older, more corroborated description, which is why a rebranded company so often finds AI still introducing it as what it used to be.

The systems drawing purely on training data make this worse, because they may have learned your old identity and will keep repeating it until they are next trained, regardless of what your live site now says. That lag is not something you can fix directly. You can only work to make the new identity so consistent and so corroborated that the next training run learns it clearly.

The practical implication is that a rebrand is not finished when your own assets are updated. It is finished when the third-party record has caught up, which takes deliberate work: updating every listing you can reach, earning fresh mentions that describe the new positioning, and giving the world consistent reasons to describe you the new way. Until then, expect AI to lag your own site, and treat closing that gap as part of the repositioning rather than an afterthought. The companies that come through a rebrand with their AI visibility intact are the ones that budgeted for the external cleanup, not just the new logo, because they understood that a model believes the accumulated web, not the announcement.

The takeaway

If AI describes your company wrongly, vaguely, or out of date, you have an entity problem, not a content shortage, and more publishing will not fix it. A model builds its picture of you from consistency across everything written about you, on your site and off it, and it recommends the entities it understands confidently. Make your name, category, and description identical everywhere you control, add correct structured data as an aid, extend that consistency to the third-party sources that describe you, and give the model specific, corroborated attributes worth attaching. Confusion costs you the recommendation; clarity earns it.

If you want to know exactly how AI currently understands your company, where the description is wrong or thin and which sources are teaching it the wrong thing, that is one of the things an AI search visibility audit measures.

This article is part of the SEO in the AI Era: The Complete Guide guide.