AI visibility for restaurants: the lever is your profile, and the number is the name
A restaurant does not get booked from a footnote. It gets booked when an assistant writes its name into the answer, and on local questions that answer is built from business profiles before it is built from websites.
AI sessions measured across the sites we track, last 30 days.
Quick answer
AI visibility for a restaurant is whether an assistant names it when someone asks where to eat and never types its name. In our own measurement pass of 3 and 4 September 2026, run on 7 domains of our travel network and not on a single restaurant, two things stood out for local questions. Google AI Mode built its answer from Google entity cards rather than websites in 34 of 41 references we checked. And the site that appeared most was cited as a source on 5 of the 5 surfaces but written into the answer text on only 2 of them.
That is the whole restaurant problem in two figures. The lever is the business profile more than the website, and the result that pays is the mention, not the link. What has been measured is how the measurement behaves on local questions, not how a restaurant fares, and the section on limits below says so plainly.
Why a restaurant is the one non hotel segment we open
Everything else on this site is about hotels, and we open restaurants for two reasons that are checkable rather than commercial. The platform already treats a restaurant as its own vertical: when the profile of a website is built, restaurant is a separate offering, and a question about where to eat that lands on a hotel site is flagged as a foreign vertical instead of being counted as a hit. That distinction exists in the code today, so the measurement does not have to pretend a restaurant is a small hotel.
The second reason is the lever itself. A hotel and a restaurant are asked for in the same shape, a place, a moment and a preference, and on that shape the surface that reads business profiles decides the answer. What we measured about local questions transfers. What we did not measure, a restaurant of any kind, does not, and we do not fill that gap with an estimate.
On local questions the lever is the profile, not the content
In a test during the pass of 3 and 4 September 2026, Google AI Mode answered local questions with 41 references and 34 of them were Google entity cards. Five out of six citations went to a business profile, not to a page anybody wrote. Read as zero website citations, that number sends a restaurant to rewrite its menu page. Read correctly, it sends it to the categories, hours, photos and reviews of its profile, which is cheaper, faster and where the citations actually are.
One reading proves nothing on its own, and that test is one surface on one pass. It is repeated weekly for that reason, and the unit cost makes repeating it trivial: Google AI Mode was the cheapest of the 5 surfaces at 0.0040 USD per question, against 0.0796 on ChatGPT. The surface that matters most to a restaurant on local questions is the one that costs least to keep watching.
Linked and not named: the gap a restaurant cannot afford
Being cited means the answer lists your website as a source. Being named means the assistant writes your restaurant into the sentence the diner reads. They come apart, and in the pass of 3 and 4 September 2026 they came apart on the best site we have: cited as a source on 5 of 5 surfaces, named in the body on 2 of them. On the other 3 the site did the work and somebody else got the recommendation.
Whether that gap hurts a restaurant more than a hotel is reasoning, not a measurement: nobody has measured how often a diner clicks after an assistant names a place. The reasoning is that a traveller often opens the sources to compare before booking, while where to eat tonight tends to be settled inside the answer, and a name that is not in the text is not on the shortlist. Treat the 5 to 2 split as the fact and this paragraph as the argument. So reporte.ai records the two things separately for every measurement, named and cited, and stores the answer text, so a zero in either column can be opened and read instead of believed. The mechanics are in prompt analysis, and what happens after a mention in AI traffic tracking.
The honest limit, and what you get inside it
We have not measured a restaurant. The pass behind this page is 57 questions on 5 surfaces over 7 domains, 285 measurements, and every one of those domains is a travel site of our own. No figure on this page describes restaurants, and any page that tells you what percentage of restaurants get named by ChatGPT without saying where the sample came from is selling you a guess.
Inside that limit the offer is concrete. The pixel, one script, so AI visits stop being filed as direct: in the 30 days to 4 September 2026 ChatGPT sent more than twenty times the visits of the next assistant across the sites we track, and in your own analytics the value to filter is utm_source equal to chatgpt.com, not openai. Then the AI readiness score, so an assistant can read what you have. Then your own questions, the ones diners in your city actually ask, measured weekly on the 5 surfaces with named and cited kept apart. The first restaurant measured will be the first restaurant figure we publish, with its sample size next to it.
No. The platform already classifies a restaurant as its own vertical in the website profile, but the pass of 3 and 4 September 2026 ran 57 questions on 5 surfaces over 7 domains of our own travel network, 285 measurements, and 0 of those domains is a restaurant. What it shows is how the measurement behaves on local questions, not how a restaurant fares.
Mostly from the business profile, on the one surface we tested for it. In a test during that pass, Google AI Mode built a local answer from 41 references and 34 of them were Google entity cards rather than websites. One reading proves nothing on its own, which is why the pass is repeated.
No, and for a restaurant the gap is the whole problem. In the pass of 3 and 4 September 2026 the site that appeared most was cited as a source on 5 of the 5 surfaces and written into the answer text on only 2 of them. A diner reads the name, not the footnote.
ChatGPT, by a wide margin. In the 30 days to 4 September 2026, across the websites we track, it sent more than twenty times the visits of the next assistant, Copilot. The live counts are in the panel at the top of this page.
The measured unit cost per question in the September 2026 pass was 0.0040 USD on Google AI Mode, 0.0129 on Perplexity, 0.0415 on Gemini, 0.0456 on Claude and 0.0796 on ChatGPT. The cheapest surface, which is also the one that reads business profiles, costs 20 times less than the most expensive.
Not first. The order is access, then the business profile, then pages, because on local questions the profile carried 34 of 41 references in our test and a page an assistant cannot read is worth nothing. Rewriting pages is the last step, not the first.
Filter on utm_source equal to chatgpt.com, not openai. In our own database chatgpt.com is the value ChatGPT actually sends, and it was more than twenty times the next assistant in the 30 days to 4 September 2026.