Your restaurant ranks first on Google and ChatGPT has never heard of it

A restaurant owner searches "best ramen in Leeds" and sees themselves at the top of Google. Then they ask ChatGPT the same thing and get four other places, or a confident paragraph about a restaurant that closed two years ago. The AI is not broken or biased. Google ranking and AI recommendation are two different retrieval systems that happen to read the same website, and they do not read it the same way. Almost every gap comes down to one of four boring causes, and each has a cheap test.

Why a restaurant can rank first on Google and still be unknown to ChatGPT

Google ranks you with a crawler that renders JavaScript, a Business Profile it trusts over your website, and years of history on your domain. It can place you first on signals ChatGPT never sees.

ChatGPT's answering crawler, OAI-SearchBot, fetches pages and does not run JavaScript. For a local query the model also grounds on a listing source. In an independent test of 2,880 prompts in August 2026, Ben Fisher found Yelp grounding in about 96 percent of ChatGPT local runs, after OpenAI's non-exclusive licensing deal with Yelp in July 2026. Foursquare appeared in none. It is one test and it will drift as deals change.

Ranking first for your own name also proves very little. Being recommended for "good Thai near the station" is a different query, and Google's AI Mode fans a question out into sub-queries you never see. Ahrefs found roughly a third of AI Overview citations come from pages ranking outside the top 100 for the parent query. The mechanics are in how AI assistants choose which businesses to recommend.

Cause one: the menu only exists in the browser

This is the first thing we check, because the site looks fine to its owner.

Across more than 500 million crawler fetches analysed by Vercel and MERJ, no major AI crawler executed JavaScript. GPTBot, ClaudeBot and PerplexityBot downloaded the script files and never ran them. Only Google's AI surfaces and Microsoft Copilot render, because they inherit Googlebot and Bingbot.

Restaurant sites are unusually exposed, because the menu is usually an ordering widget, the hours load from a script and the address sits inside a map embed. Take the JavaScript away and the page is a logo, a hero photo and a "Book now" button that goes nowhere.

The test takes two minutes. Open the site with JavaScript disabled, or fetch the raw HTML with curl and read it. If the dishes, prices, hours and street address are not there, ChatGPT, Claude and Perplexity have never seen them. Google has, which is why you rank.

The fix is plain HTML for the menu, hours and address, even if a widget also loads on top. A PDF menu is worse than HTML and better than a widget; keep it small, because ChatGPT's crawler rejects pages over 4 MB.

Cause two: your CDN is blocking the crawler you need

The second cause lives in a dashboard nobody at the restaurant has opened since the site launched.

Most CDNs and hosting panels now offer a single switch that blocks AI crawlers as a group. Bot protection that serves a JavaScript challenge does the same thing by accident, because a crawler that cannot run JavaScript cannot pass the challenge. From the crawler's side it looks like a 403 or an empty page. Open the CDN firewall log, filter for OAI-SearchBot, and look for challenges or blocks. If they are there, that is the whole diagnosis.

There is a quieter version in robots.txt. GPTBot is the training crawler. OAI-SearchBot is the answering crawler, and blocking it removes you from ChatGPT search entirely. Anthropic has the same split: ClaudeBot trains, Claude-SearchBot answers. Google-Extended, which people also block, has no effect on AI Overviews or AI Mode. The full token list is in our AI crawlers and robots.txt guide.

Check the 404s too. Vercel and MERJ measured 404 rates of about 34.8 percent for ChatGPT's crawlers and 34.2 percent for Claude's, against 8.2 percent for Googlebot. If your menu moved URL two redesigns ago, Google worked it out and ChatGPT may still be requesting the old address.

Cause three: your listings disagree with each other

Even with a readable page and an open door, the assistant has to decide whether to trust what it read. It does that by checking other sources. For a restaurant those are the listing sites, and for ChatGPT local answers, mostly Yelp.

The website spells the name with an ampersand and Google Business Profile spells it with "and". Yelp shows the old address and the previous owner's hours because nobody claimed the listing. To a model reading all of that, you are three weak entities instead of one strong one, and it names the place down the road whose listings agree.

The test is tedious. Open your website, Google Business Profile, Yelp, TripAdvisor, Apple Maps and the ordering platform side by side. Compare name, address, phone, hours and cuisine label. Every difference is a reason for a model to hedge, and hedging means naming someone else. Then claim and correct your Yelp listing.

Cause four: nobody independent has written about you

This is not a technical fix, and it is what separates ranking from being recommended.

Ahrefs looked at 75,000 brands and found branded web mentions correlated with AI Overview visibility at 0.664. Backlinks correlated at 0.218. Muck Rack analysed more than 25 million citations and found earned media accounted for 84 percent of them, and paid or advertorial placements 0.3 percent. Both are correlations. Neither proves cause, and we would not sell them as more than that. But the direction is the same in two unrelated datasets, and it points at the local paper's "where to eat" piece and the neighbourhood roundups you did not write.

One caveat cuts the other way. Semrush's ghost citation study found 61.7 percent of AI citations produced no brand mention. ChatGPT cited a source 87 percent of the time and named a brand 20.7 percent of the time. Being in a roundup is necessary and still not sufficient. The roundup has to name you clearly, near the top, with a reason.

The diagnostic in order

Run these in sequence and stop at the first failure, because each one masks the ones below it.

Step five exists because a Search Engine Land instrumented study found ChatGPT stores a snippet of roughly 200 characters per page, taken from the text near the H1, query-independent and frozen at index time. "Welcome to our website" wastes it. The cuisine, the street and the nights you are open do not.

How to tell whether any of this worked

Not from traffic. Conductor's benchmark puts AI referrals at about 1 percent, many arrive with no referrer and land as Direct, and Google Search Console's Generative AI report shows impressions only.

Two things do work. Bing Webmaster Tools has an AI Performance report, launched in February 2026, that shows Copilot citations and the grounding queries the model generated. And you can sample prompts yourself, provided you repeat each one at least 12 times, because model outputs are non-deterministic and five repetitions carries a margin of error of roughly plus or minus 27 points. Appearing twice in five runs tells you nothing. The setup is in our guide to tracking AI search visibility.

Start with the menu. Disable JavaScript, load the page, and see whether your own dishes are still there. It is the check most owners have never run, and the only one on this list that costs nothing to fix.