Before Visibility, Identity: Does Your Brand Exist as an Entity?

Most brands arrive at AI visibility with the wrong first question. They want to know whether ChatGPT recommends them. Fair enough, but there is a question underneath it that has to be settled first, and almost nobody asks it: does the model know you exist at all, as a specific thing, distinct from everything else that happens to share your name?

If the answer is no, every metric you collect afterwards is measuring something else. You cannot be recommended if the machine cannot work out who you are.

What an entity is, in plain terms

To a language model, your brand name is a string of characters that appears in a certain amount of text. In the training data it sits next to other words: products, industries, cities, founders, competitors, complaints, reviews. What the model learns is the shape of that neighbourhood.

An entity is what happens when the neighbourhood is coherent enough to be treated as one thing. The model, and the retrieval systems bolted onto it, can then attach facts to a stable referent: this name is a company, it is in this category, it sells this, it is not the other one. That is entity resolution, and it is the difference between a name a machine can reason about and a name it can only echo back at you. A list of links can hedge and show both meanings of an ambiguous name. An answer written in prose has to pick one.

The three ways a name fails to resolve

Your name is also an ordinary word

Call your company Meridian, or Compass, or Orbit, and you have handed the model a word that already carries meaning. Ask about the brand and it may define the noun, or blend the noun into the company, or describe a beautifully generic organisation assembled from what such a word implies. Common-word names are excellent for humans, who use context effortlessly, and hostile to machines, which must infer it.

Somebody else already has your name

Two companies sharing a name, in different countries or industries, is the normal state of affairs. If one is older, larger, or simply better documented online, the model resolves the name to that one and quietly attributes your mentions to it. Their category becomes yours. Their bad press becomes yours.

The name belongs to a person

Firms named after a founder hit this constantly. The surname also belongs to an athlete, a politician, an author, and the web talks about that person far more than about your business. The model is not confused, exactly. It has learned that this string usually refers to somebody else.

The consequence is the same in all three cases: your mentions disperse. What you build never accumulates onto one thing. It scatters across several possible referents, and none of them is quite you.

Where entities actually get defined

Models do not take your word for who you are. They take the web's word, weighted by how independent and how consistent the web sounds. Some places carry more weight in that process, not because anyone declared them official, but because of what they are: neutral, structured, written by people with no commercial stake in you. Collaborative encyclopedias such as Wikipedia are the obvious example, along with the structured knowledge bases behind them, such as Wikidata, where an entity is not a paragraph but a record with identifiers and typed relationships. Alongside them sit business registers, industry registries, standards bodies, professional associations and sector directories.

What makes a brand resolvable is a definition that is consistent and repeated across independent places. When several unrelated sources describe the same name the same way, in the same category, with the same founding facts, the ambiguity collapses: there is enough agreement for a machine to commit. When each source says something slightly different, or nobody says anything, there is nothing to commit to. That is why an encyclopedic entry, where one legitimately exists, has an outsized effect. Not because models revere the site, but because a well-sourced entry is a compact, neutral, heavily mirrored statement of what a name refers to, and other datasets inherit it.

The part you will not like

You do not write your own encyclopedia entry. You do not pay an agency to write it either, and you do not pay a friendly editor to slip it past review. Those communities have explicit rules on conflict of interest, notability and independent sourcing. An entry about a company that is not the subject of substantial coverage in sources unconnected to it gets deleted, and a promotional entry gets deleted faster. The failed attempt is worse than the absence: it leaves a record, on the open web where the models read, of a brand caught fabricating a fact about itself.

If you do not meet the notability criteria, the honest reading is that you are not notable by those criteria, not yet. That is not a verdict on your business: plenty of profitable, well-run companies never qualify. It means the encyclopedia is not your lever, and you can stop staring at it.

What you can legitimately do

• Be ruthless about the name. One spelling, one capitalisation, one legal suffix or none at all, everywhere: site, profiles, invoices, press releases, partners' pages. Every variant you tolerate is another referent a machine must reconcile.
• Put structured data on your own site. Publish an Organization record in JSON-LD with your legal name, category, logo, founding details, and above all the sameAs property pointing at your profiles and registry entries elsewhere. That property exists to assert that the scattered pages are one entity, and the entity is you.
• Write an about page a machine could quote. Not a manifesto. One paragraph near the top saying what you are, what you sell, who buys it and which category you sit in, in the words your market uses. If a model must infer your category from positioning poetry, it will infer something, and you will not enjoy it.
• Claim the places that accept everyone. Business registers, trade associations, sector directories, developer registries, app marketplaces, the partner pages of tools you integrate with. None demand notability, and each repeats the same definition from an independent address.

None of this guarantees an outcome. It removes an obstacle. The two are not the same, and any vendor blurring them is selling something they cannot deliver.

The test, which takes a minute

Open ChatGPT, Claude, Gemini and Perplexity and ask each one the flattest possible question: who is [your brand]. You are not grading the praise. You are checking whether the machine has you confused with someone else.

Watch for the tells. A category that is not yours. A country you have never operated in. A founder who is not your founder. A fluent description of a different company with your name. A polite refusal to say anything, which is its own kind of answer. Then ask again in every language you sell in: a name that resolves cleanly in English can collide with something ordinary in German or Spanish, and the model will pick the local meaning.

If the answers come back confused, your problem is not visibility. It is identity, and no amount of content will fix it until the machines know which thing you are.

Where measurement fits

Once the name resolves, the useful question becomes which sources the models lean on when they describe you, and whether that description survives the crossing into your export markets. That is measurable, and it is what PSentry does: it runs your prompts across ChatGPT, Claude, Gemini and Perplexity, per language and per market, and reports where you are mentioned, which sources were used, and who gets named instead of you. The sources matter here: they tell you which pages are defining you, and whether those pages describe the right company.

What it will not do is fix the resolution problem for you: it cannot write your encyclopedia entry, it cannot edit what a model has already learned about your name, and it cannot promise that a confused answer becomes a clean one. What it tells you is whether the machines currently know who you are, market by market: the fact this discipline rests on, and the one most brands have never checked.

Frequently Asked Questions

Should we rename the company if the name is ambiguous?

Almost never. Renaming destroys the recognition you have with humans, and humans sign the contracts. Described consistently in enough independent places, your meaning becomes the dominant one inside your category, which is the cheaper fix.

Does a Wikipedia entry guarantee that AI systems describe us correctly?

No. It helps a model attach facts to the right referent, because the entry is neutral, structured and widely mirrored. It does not make you recommended, and an entry full of claims nobody else corroborates is fragile: it can be cut there, and the models weight agreement across sources anyway.

We are young and have no press coverage. Is entity work pointless for now?

No, just narrower. Name consistency, structured data and an unambiguous about page cost nothing and sit in your hands. The parts that depend on independent coverage are earned later. Do the controllable half now, and the coverage, when it comes, lands on an entity that already exists.

An assistant describes a competitor when we ask about us. What now?

Work out whether it is a real competitor or a company that merely shares your name: the remedies differ completely. A namesake is a resolution problem, answered by consistency and independent corroboration. A rival named in your place is a visibility problem, and it means your entity is fine.