There is a question that shortcuts most of the confusion about AI visibility, and almost nobody asks it. Not how do I rank in ChatGPT — there is no ranking. The useful question is: when an assistant recommends a business in my category, where did that name come from?
Where an assistant shows its citations, this stops being theoretical. You can read the list.
The pages that keep appearing
Do this across enough categories and cities and the same shapes recur. Not the same websites — those differ by industry and country — but the same kinds of websites.
Directories and listings. The industry-specific ones matter more than the general ones. A trade association’s member list, a professional register, a category-specific platform. These are dense with exactly the information an assistant needs: name, category, location, sometimes credentials, all in a consistent format across many businesses.
Roundups and “best of” articles. Someone wrote The 8 Best Physiotherapists in Bristol in 2023. It is possibly out of date, possibly written by someone who never visited any of them, and it is one of the most influential pages in that market, because its structure is precisely the structure of the answer the assistant is being asked to produce.
Review platforms. Not only for the star ratings, but because review pages carry a great deal of natural language about a specific business — the phrases customers actually use, attached to a name. That is unusually rich material for a model.
Forum and community threads. A discussion where someone asked your exact question and got five candid replies is, from a model’s point of view, very high quality: real people, weighing real options, giving reasons.
Local press and institutional pages. Chamber of commerce membership, a local news piece, a supplier’s “where to find us” page. Individually minor, collectively a strong signal that a business is real and locatable.
- The customer asks in their own words Which clinic should I use in Tijuana — a question about a problem, not about any business by name.
- The assistant rewrites it as a search One question becomes one or more queries, phrased the way the search back-end expects rather than the way the customer typed it.
- It opens a handful of pages Often three or four, not ten. Pages that already compare several options win here, because they match the shape of the answer being requested.
- It lifts the names and the reasons Whatever those pages say about who is good at what becomes the answer, in the assistant’s voice and without the hedging the source may have carried.
For “which one should I pick”, the names are chosen at the third step and your site enters at the last, as confirmation. For concrete questions, your site can be one of the pages opened at the third.
A measured example: restaurants in Mexico City
On 23 September 2026 we asked ChatGPT five questions about where to eat in Mexico City and recorded every source it cited. There were 21 citations:
| Type of source | Citations | Which |
|---|---|---|
| Guides and publications | 12 | El País (5), Michelin Guide (4), Eater, travel guides |
| Platforms | 6 | OpenTable (5), Discover CDMX |
| Restaurants’ own sites | 3 | Rosetta, Contramar, Nicos |
| Google Maps and TripAdvisor | 0 | — |
The “best of” articles and the press did most of the work, as the section above describes. And review platforms, which that same section lists as a frequent source, did not appear once. Sources change by category and by city, so they have to be checked for your case rather than assumed.
It is a small sample: five answers, one assistant, one day. The detail is in how to get ChatGPT to recommend your restaurant.
Where your own site counts, and where it doesn’t
It depends on the question. When someone asks “which clinic should I use in Tijuana?”, the assistant needs several names with distinguishing reasons. A single business’s website supplies one name, described in promotional language that fits every competitor equally well. For those questions, your site is rarely the reason you were named.
Concrete questions are different. In our index of 125 ChatGPT answers across five Mexican cities, when someone asked which hotel has parking, 14 of the 20 sources were the hotels’ own sites. Asked where to get a cheap breakfast, ChatGPT cited cafés’ menus and repeated their prices. Asked about bar promotions, it copied the ones published with day, time and price. Across the whole index, businesses’ own sites were 20% of citations.
The practical rule: your site will not put you on the “best of” list, but it can win the concrete question you are genuinely the best answer to, as long as the fact is written there, as text. And once a third party has named you, it is where the assistant confirms what you do, where, for whom and at what price.
The practical consequence
Because the influential sources are a short list rather than the whole web, this becomes an addressable problem rather than a vague one.
For any given category and city, the citations tend to concentrate: a handful of sources drive most of the answers. Once you know which, you can work through them one at a time:
- Where you’re listed but wrong. Extremely common, and the cheapest fix available. An old address, a service you stopped offering, a phone number from two premises ago. You already occupy the position; the information in it is just stale.
- Where you’re absent but eligible. An association you belong to but never completed the profile for. A register that requires only an application.
- Where a competitor is and you aren’t. A roundup with a named author who can be contacted. A review platform where you have a neglected profile.
- Where you cannot realistically appear. Some sources are closed, paid, or editorially hostile. Knowing this is still worth something: it tells you where not to spend effort.
That sequence is worth more than any amount of rewriting your homepage, and it is available to any business willing to look.
The catch
Citations are not the whole story. When an assistant answers from training data rather than live search, it cites nothing — and the influences are older, broader, and not directly visible. You cannot read them; you can only infer them from what the answer says.
This is why an answer with no citations should not be read as an answer with no sources. It usually means the sources are old enough to have been absorbed during training, which is both harder to change and more durable once changed.
It is also why a single check tells you very little. The same question asked twice can produce one cited answer and one uncited one, naming different businesses. Understanding which sources genuinely control your category takes repetition, across assistants, across phrasings — which is a measurement problem, not a browsing problem.