How does ChatGPT choose which businesses to mention?

Any confident, detailed answer to this question should be treated with suspicion. The exact selection behaviour of ChatGPT and comparable assistants is not fully public, differs between models and versions, and changes without announcement. Nobody outside those companies can describe the mechanism precisely, and the people who claim to usually want to sell you something.

What can be described honestly is the general shape of the process, what it plausibly depends on, and where the genuine uncertainty sits. That is enough to act on.

Two different sources of knowledge

Assistants draw on roughly two things. The first is what the model absorbed during training — a large body of text with a cut-off date, containing whatever was publicly written about businesses up to that point. The second is what the system retrieves at the moment you ask, where the product has browsing or search available.

The distinction matters. Information that exists only in training data can be outdated and cannot be corrected quickly. Information that can be retrieved live reflects the current web but depends on your content being findable and fetchable right now. A business that is invisible to retrieval is relying entirely on what a model happened to absorb months or years ago.

Relevance to the actual question

Selection starts with matching, and the match is semantic rather than literal. A question about a leaking roof after a storm is not looking for pages containing the phrase “leaking roof”; it is looking for businesses that plausibly handle emergency roof repair in a particular place.

This is why describing your services in customer language matters more than keyword coverage. A business that lists concrete services and problems it solves can be matched to far more questions than one describing itself in abstractions.

Clarity, because ambiguity is expensive

A system generating an answer is exposed if it says something wrong. That pressure pushes it toward businesses it can describe unambiguously. If it cannot tell whether two records describe the same company, whether you still operate, or which of two addresses is current, the cheapest move is to write about a business where none of those doubts exist.

Evidence and corroboration

Claims that appear in more than one independent place are safer to repeat than claims that appear only on the business's own site. Registration records, professional bodies, review platforms, local press, and directories all serve as corroboration.

This is not a reputation score, and it is not a count of mentions. It is closer to the ordinary editorial instinct of preferring a fact you can confirm twice.

Context, geography, and intent

The same words mean different things depending on what surrounds them. “Best studio near me” from someone who previously asked about wedding photography is a different question from the same words after a conversation about recording music. Assistants use the conversation, and sometimes location, to disambiguate.

Intent shapes selection too. “Who should I call today” favours businesses with clear availability and contact details. “Who is best for a complex commercial project” favours demonstrated depth. A business that only ever describes what it is, never how it works or who it suits, is hard to attach to any specific intent.

What is genuinely uncertain

Several things deserve explicit acknowledgement rather than confident speculation:

• How much weight is placed on training data versus live retrieval, for any given product and question.
• How sources are prioritised when they conflict, and what tips the balance.
• Whether a specific business appears because of one strong signal or many weak ones.
• How much randomness is involved — the same question can produce different businesses on different days.
• How much any of this will resemble itself a year from now.

That last point is the important one. Any strategy built on the current behaviour of one product is fragile. A strategy built on being accurate, consistent, and genuinely informative survives changes to the mechanism, because every plausible version of the mechanism rewards those things.

What this implies in practice

Make your business easy to identify and hard to confuse with anyone else. Describe services and service areas concretely. Keep public information consistent so there is nothing to reconcile. Make sure your content can actually be fetched. Support claims with something checkable outside your own site. Publish real answers to real questions.

None of that guarantees inclusion, and you should be wary of anyone who says otherwise. It changes the odds by removing the reasons you are currently being passed over.

For the specific facts to get right first, see what information AI needs about a company. For why search practice alone is not sufficient, see is traditional SEO enough for AI visibility. Our process and services start by recording what assistants currently say about you, so the guesswork is replaced with observation.