How restaurants get recommended by ChatGPT

ChatGPT recommends restaurants it can describe confidently, from sources that agree with each other. It is not judging your food — it cannot taste anything — and it is not ranking you the way Google does. It is assembling a short answer from what it knows, and it favours businesses whose description is unambiguous, specific and repeated across several credible places.

Which means the question "how do I get recommended by ChatGPT?" has a fairly unromantic answer: give it one clear, distinctive, factual description of your restaurant, and make everything on the internet agree with it.

Where does ChatGPT get its restaurant information?

Two routes, and they behave differently.

When it browses, which it does for anything local or recent, it runs searches and reads the pages it finds. In practice that means your website, your Google Business Profile as surfaced in search results, TripAdvisor, local food blogs, regional publications and any listing site ranking for your town. Whatever ranks well in normal search is disproportionately likely to be read.

When it does not browse, it answers from training data — everything it absorbed about your town before its cutoff. This favours restaurants that have been described in many places over a long period, which is harder to influence quickly and is why the browsing route is where a small restaurant's effort pays back fastest.

Either way, the inputs are ordinary web sources. There is no restaurant database, no partnership, no submission process.

Why does it name the same few restaurants repeatedly?

Because those restaurants are easy to summarise, and a model writing three sentences needs to be able to summarise.

Picture the model's problem. Someone asks for dinner recommendations in your town. It has maybe fifteen candidate restaurants and room for four. Restaurant A is described as "Newari cuisine, forty covers, near the lake, moderately priced" by its own site, its Google profile, TripAdvisor and a travel blog — four sources, no contradictions. Restaurant B calls itself "a fine dining experience" on its site, "Asian, vegetarian friendly" on TripAdvisor, and "Nepalese restaurant" with no detail on Google.

Restaurant A gets named because the model can write a sentence about it that will not be wrong. Restaurant B gets skipped, not because it is worse, but because summarising it accurately requires a judgement call the model would rather avoid.

I watched this play out with a 40-cover restaurant in Pokhara that was named in zero of twenty AI answers despite a 4.6 rating and 210 reviews. Three competitors with fewer reviews appeared repeatedly. The only meaningful difference was that they described themselves the same way everywhere.

What actually makes a restaurant citable?

Five things, roughly in order of how much they matter for an independent restaurant:

1. One distinctive factual claim. Something true, specific, and not shared by four hundred other businesses. Cuisine plus a detail: "traditional Newari, wood-fired, forty covers." "Great food in a warm atmosphere" is not a claim, it is wallpaper.

2. Agreement across sources. Your site, Google Business Profile, TripAdvisor, Agoda, delivery platforms, directories. Same cuisine classification, same price range, same hours, same description. This is the single highest-leverage item and it costs nothing but an afternoon.

3. A machine-readable menu. A photograph of a printed menu is invisible — to Google and to any model reading your page. A real HTML menu with Menu and Restaurant schema, listing dishes, descriptions and prices, gives a model the exact detail it needs to say "known for its bara and choila, around NPR 800 a head."

4. Reviews that say something. Volume and rating help, but specificity helps more. Reviews that repeatedly mention a particular dish, a particular occasion ("good for groups", "quiet enough to talk") or a particular trait give a model concrete language to reuse. You cannot write reviews, but you can ask well: "if you enjoyed the momos, mentioning them helps other people find us" produces measurably more specific reviews than "please leave us a review."

5. Third-party corroboration. A food blogger, a regional publication, a well-regarded local guide. One good mention outperforms fifty directory listings, and it is what lets a model treat your self-description as verified rather than claimed.

Does answering questions on your website help?

Yes, and it is the most underrated item on the list. Pages that answer a specific question in the first sentence get quoted far more often than pages that build up to a point.

The mechanism is straightforward: a model extracting an answer takes the opening of a relevant passage. If your page about your cuisine opens with "Our chef has been perfecting his craft for over twenty years, drawing on family recipes passed down through generations", there is nothing in that sentence to extract. If it opens with "Newari cuisine is the food of the Kathmandu Valley's indigenous community, built around buffalo, beaten rice and fermented greens; the dishes most visitors start with are bara, choila and yomari", you have handed it a quotable paragraph.

On the Pokhara restaurant, a 900-word page answering "what is Newari food and what should I order first?" became the cited source in seven of eleven answers where the restaurant was named. One page, written to be lifted, outperformed everything else on the site.

Write self-contained paragraphs, too. Each one should make sense pulled out of context, because that is exactly how it will be used. A paragraph beginning "This is why we do it differently" is useless on its own.

How do you know if any of this is working?

You test it monthly, on a schedule, with the same queries. There is no dashboard, no notification and no analytics event when a model names you.

Pick twenty queries a real guest would type — "where to eat authentic [cuisine] in [town]", "best dinner near [landmark]", "restaurant in [town] for a group of eight", "good vegetarian food in [town]" — and run each through ChatGPT, Perplexity, Gemini and Google AI Overviews. Log every business named, the date, and which source was cited where one is shown. Ninety minutes a month.

Expect it to be uneven. Perplexity and AI Overviews respond fastest because they lean hardest on live search. Gemini and ChatGPT move more slowly. A win in one does not transfer to the others, which is why you track all four rather than checking your favourite once and drawing conclusions.

What does not work

• Paying for "AI directory" listings. These cold emails are constant and the product does not exist. Nobody sells entry into a model's answer.
• Stuffing your pages with "best restaurant in [town]". Models are not matching keywords, they are extracting facts. Keyword stuffing removes the facts and adds nothing.
• Publishing twenty thin blog posts. One page that genuinely answers a question outperforms twenty that circle one.
• Fake reviews. Beyond being against every platform's terms, they produce exactly the generic, non-specific language that makes a restaurant harder to quote, not easier.

Where to start this week

Run the baseline. Twenty queries, four assistants, one spreadsheet. Then open your Google Business Profile, TripAdvisor listing and website side by side and read the descriptions in sequence. Most restaurant owners doing this for the first time find at least two contradictions within five minutes — a cuisine type that no longer applies, an old price range, hours that changed last season.

Fix those, structure the menu, and write one page that properly answers the question your guests ask most. That is the first month, and it is enough to tell whether the rest is worth your time. If you would rather not run it yourself, it is what my GEO and AEO work covers, starting with the same baseline log.