Where ChatGPT gets its restaurant recommendations

Ask ChatGPT for a good ramen place in Manchester and it names a handful. That list comes from two places: licensed listing and review data, which since July 2026 means Yelp, and web pages fetched live by OAI-SearchBot. An independent test found Yelp behind roughly 96 percent of local runs. Your website is the smaller half of the answer, and Foursquare is no part of it.

Where ChatGPT gets its restaurant recommendations

ChatGPT draws on three things when it answers a local query.

The first is licensed data. OpenAI signs deals with companies that hold structured listing and review data, and the model is handed that data at answer time. Since July 2026 the partner that matters for restaurants is Yelp, under a non-exclusive licensing deal.

The second is live web retrieval. OAI-SearchBot, the answering crawler for ChatGPT search, fetches pages from OpenAI's own index and from third-party search providers. Your website, a local paper's "best brunch" round-up and a food writer's review all reach ChatGPT this way, and only if the page can be read without JavaScript. The Vercel and MERJ analysis of more than 500 million crawler fetches found no major AI crawler executing JavaScript, so a menu that renders in the browser is invisible to it.

The third is the model's training memory, which is old, unattributed and cannot be edited. It is the weakest of the three for anything with opening hours.

So the honest picture is licensed listings plus web results, weighted heavily towards the licensed side for local queries.

What the Yelp deal changed

In July 2026 OpenAI signed a non-exclusive licensing deal with Yelp. Non-exclusive matters. It means Yelp can license the same data to other AI providers, and it means OpenAI can add other partners later. Neither has been shown to be happening for restaurant answers yet.

In August 2026 Ben Fisher published an independent test. He ran 2,880 local prompts through ChatGPT and recorded which sources the model grounded on. Yelp appeared as a grounding source in roughly 96 percent of runs.

That is a very high number for a single source. When ChatGPT answers a restaurant question, it is very likely reading your Yelp listing before it reads anything else. Your category, your hours, your address, your price range and your review text are the inputs. If they are wrong on Yelp, they are wrong in the answer.

Why the Foursquare story is unsupported

A story has circulated that Foursquare data powers ChatGPT's local results. It gets repeated often enough that many owners take it as settled.

Fisher's test measured it. Across the same 2,880 prompts, Foursquare citations in ChatGPT local results came to zero percent. That is as close to a clean negative as a black-box test gets.

Two caveats, because this is a black box. The test is a snapshot from August 2026, and a non-exclusive deal leaves the door open for OpenAI to add a second data partner. If that happens, someone will need to run the test again. Until they do, the Foursquare story has no evidence behind it and a fairly large test against it.

Do not pay anyone to "optimise your Foursquare presence for ChatGPT". Keep that listing accurate as a matter of hygiene, then spend the hour on Yelp.

What this means for a cafe or a hotel

The order of operations changes. Most GEO advice starts with your website. For a local hospitality business the evidence says start with the listing and review surfaces, because that is what the model reads first.

Here is the order we would work in.

Yelp first

Claim the listing if it is unclaimed. Check the category is the most specific one available, not a broad parent. Confirm the hours match the door. Add photos of the room and the food. Reply to reviews, because the review text is part of what the model reads.

For a hotel the fields are the same and the retrieval path is the same. We have not seen a published test that isolates hotel prompts, so treat Yelp-first as a well-supported default rather than a measured result.

Then Google, for Google's surfaces

Google's AI Overviews and AI Mode do not run on the Yelp deal. They run on Google's own index and on Google Business Profile. Fix the same fields there. There is a separate guide to Google AI Overviews for local business if you want the detail.

Then your own site, and the paragraph after the H1

Your site is still the second half. The Search Engine Land instrumented study found ChatGPT stores a snippet of roughly 200 characters per page, query-independent, frozen at index time, taken from the text near the H1. That is the extract ChatGPT holds for you regardless of what someone asked. Make it say what the place is, where it is and the one thing it is known for, in plain words.

Confirm OAI-SearchBot is not blocked in robots.txt. It is the answering crawler. GPTBot is the training crawler, and blocking it does not remove you from ChatGPT search. The AI crawlers and robots.txt guide has the full list.

And fix broken links. Vercel and MERJ measured 404 rates around 34.8 percent for ChatGPT's crawlers against about 8.2 percent for Googlebot. A stale menu URL is a fetch that returns nothing.

Then mentions on sites you do not own

Ahrefs studied 75,000 brands and found branded web mentions correlated with AI Overview visibility at 0.664, against 0.218 for backlinks. Muck Rack, across more than 25 million citations, found earned media accounted for 84 percent of AI citations and paid or advertorial placements for 0.3 percent. Both are correlational. Both point the same way. A mention in the local paper's brunch list does more than another page on your own site.

Being retrieved is not being named

Semrush's ghost citations study found 61.7 percent of citations produced no brand mention at all. ChatGPT cited a source in 87 percent of answers and named a brand in 20.7 percent.

So ChatGPT can read your Yelp listing, read your site, cite a round-up that includes you, and still not say your name. What tends to earn the name is the specific claim: the thing you are known for, stated the same way on Yelp, on your site and in the press piece. If the three surfaces describe you differently, the model has less reason to pick you out.

How to check what ChatGPT is reading about you

Do not trust a single run. Outputs vary from run to run. Five repetitions of a prompt carry a margin of error near plus or minus 27 points, and a usable sample needs 12 or more. Pick a handful of prompts a real customer would type, run each at least 12 times, and log which sources appear. Yelp will probably dominate. Note whether your site appears at all, and whether any third-party page does.

Then accept that you will not see the result in analytics. AI referrals are a small share of traffic, around 1 percent in Conductor's benchmark, and many arrive with no referrer and land as Direct. The guide to tracking AI search visibility covers what can be measured and what cannot.

The broader playbook is in how restaurants get recommended by AI. This post is the narrower point underneath it: for ChatGPT specifically, the listing you have probably not opened since you claimed it is the document the model reads first. Open it today and check the hours.