Recommended by AI in 90 days

Your restaurant in AI answers: a menu that is a picture, so the allergens and the late seating never get quoted

A diner in Providence asks ChatGPT which restaurant serves dinner after 10 on a weeknight and lists the allergens in the menu, not in a PDF, and it names two places from a tourist roundup while the room that seats until 10 is absent. Here is where those names come from, the picture that hides the dish, and a hypothetical worked example.

Restaurant staff prepare a dining room
MemberVIP case · 90 days
44%of sign-ups came from ChatGPT, more than from Google.
See the MemberVIP case →

Free test

See what AI answers today when your customers ask.

Free, no card. We save the answers as your starting point.

The question a diner actually asks

Owners usually meet this as a full Saturday, not as a lost table. They ask an assistant about their own restaurant and it praises the room, because the question already contained the name. A stranger then asks the way a person with an allergy asks before they book: what is in the dish, and whether the kitchen is still seating. The restaurant is gone. The only reading that matters is that stranger’s question, asked the same way every week.

Almost nobody types “best restaurant near me” and stops. They add the constraint that decides the night. A table after 10. A dish without gluten, or a list of what the pasta does contain. A photographed menu is not the criterion. The assistant cannot read a picture of a chalkboard, and a PDF from March is a file, not a sentence it can quote. It can quote an HTML line that names the dish, the price, and the allergen.

Where the answer about a restaurant comes from

With web search on, each assistant runs a few searches and summarizes what the pages agree on. For a dining room the sources are Google, review sites, the local paper’s lists, delivery apps, and the restaurant’s own pages. A roundup ranks for a broad question because many pages repeat the same five names. A small room ranks for the late-seating question only if a page states the last seating in a sentence, and for the allergy question only if the menu is text.

The mechanism is a sentence a stranger can hold the kitchen to. A photo of a plate does not say what is in it, and “ask your server” is not a list. The four assistants disagree because each searches a different index, and the same one changes between runs, so one screenshot proves little. Fame is not the criterion: the place on three “best of” lists wins “restaurant in [city]”, and the allergen question goes to whoever printed the dish as text.

The most common mistake: the menu is a picture

Almost every independent restaurant publishes the menu as a PDF or a photo, and the hours as “dinner” with no last seating. It happens because the designer’s file is what the printer uses, and updating a PDF feels like the whole job. To an assistant, a picture and a PDF are the same absence. It cannot quote “sesame” from a photograph, and it will not guess a closing time from “open late”.

The cost is concrete. The diner with an allergy takes the room whose page lists the allergen next to the dish. The better kitchen may cook that dish every night, and it never enters the answer. The remedy is an HTML menu: each dish, the price, and the allergens in it, plus one sentence for the last seating. The trade-off is that a text menu is less pretty than the photograph, and it has to change when the dish changes. A pretty file that cannot be quoted does not fill the table.

A worked example, step by step

Example: a hypothetical restaurant, not a sayurname client. Imagine Fox Point Table in Providence, Rhode Island: the chef and fourteen people, 46 seats, eight years, 210 Google reviews averaging 4.6. Last seating is 10:00 Tuesday to Thursday, kitchen at 10:30, closed Monday. Pasta is $19 to $28. The menu names, next to each dish, milk, egg, fish, shellfish, tree nuts, peanuts, wheat, soy and sesame when they are in it. The old site showed a PDF from March and a photo of the chalkboard. Hours said “dinner”.

The measured question was “Which restaurant in Providence serves dinner after 10 on a weeknight and lists the allergens in the menu, not in a PDF?”. At week 0 the room was named by 0 of 4. ChatGPT named two places from a tourist roundup. Gemini listed the highest-rated rooms on Maps and their star counts. Perplexity cited a 2024 “best Italian” list. Claude described how to ask a server about allergens and named nobody.

  1. Read the four answers with the chef and chose ten questions diners had asked that year: a table after 10, the allergen in a pasta, a gluten-free dish, a quiet corner, a birthday, and what the kitchen does not serve.
  2. Fixed the base: the same legal name, address and hours on the site and on Google, including the Monday close.
  3. Replaced the PDF with an HTML menu: each dish, the price, and the allergens in that dish, written as text.
  4. Wrote the hours as a sentence: last seating 10:00 Tuesday to Thursday, kitchen at 10:30, closed Monday. “Open late” came off the page.
  5. Removed “the best Italian in Providence”. It moved nothing, and no source can confirm a superlative.
  6. Kept a monthly note when a dish changed, each read by the chef, and asked guests to mention the allergen list in a review, with no discount and no suggested wording.

Nothing moved for five weeks. In week 6 Gemini named the room and quoted the 10:00 last seating. In week 10 Perplexity followed, after two reviews said the allergen was on the menu and not a surprise at the table. At week 12 ChatGPT still named the tourist roundup, and Claude named the room in one run of four, which does not count.

What did not work: a paid “best of Providence” badge brought clicks and no change in any answer. The brunch question never moved, because the kitchen does not serve brunch and two cafés publish that menu in text. Results depend on the city, the question and the competition. Another room might see none of this in twelve weeks.

Diagram of the hypothetical example: the measured late-seating and allergen question, the six steps as numbered boxes, and a grid showing the restaurant named by none of the four AIs at week 0 and by Gemini and Perplexity at week 12.
Hypothetical example: Fox Point Table is an illustration, not a sayurname client.

What to publish

A citable restaurant page answers one constraint with lines a stranger could book against. Titles that work: “Last seating at Fox Point Table: 10:00 Tuesday to Thursday”; “Allergens on the menu, dish by dish, not in a PDF”; “What we do not serve: no brunch, no delivery-only menu”.

An excerpt: “Last seating is 10:00 Tuesday to Thursday. The kitchen closes at 10:30. We are closed Monday. Each dish on the menu names milk, egg, fish, shellfish, tree nuts, peanuts, wheat, soy or sesame when that dish contains it. The menu is text, not a PDF.” What makes it citable: a clock time, a day, a list, and a format an assistant can quote.

When it is not worth it

If the room is full Tuesday to Thursday and the only empty seats are at 5, more late-seating inquiries do not help. Change the hours or add a turn before you publish a promise you cannot seat. A delivery-only kitchen with no room will not be chosen by someone booking a table.

Nor does it pay when the allergen list is a guess, or the PDF is the only menu the printer will honor and nobody will keep the text in step with it. Start there. Come back when the sentence on the page is a seating you will actually take.

How the work runs

It starts with the free test: one diner’s question sent to the four AIs, with the full answers saved as the starting point. Once subscribed, we measure up to 10 questions every week in ChatGPT, Gemini, Perplexity and Claude, fix the site’s base so the assistants can read the menu as text, and publish as much as we judge necessary for AI to cite your business in the answers you want. Volume and format follow what each question needs, and that part is our call. We also ask for inclusion in the sites the assistants use to answer each question. What we promise is the request, not the listing: many sources are editorial or paid and do not accept one, and you see what was requested and what got in. Publishing happens on a subdomain connected by DNS, or we deliver the pieces and you publish them. Onboarding is online, with no call.

It costs $299/month, with no lock-in. A one-time $599 bonus is charged only when an AI starts naming your restaurant for a measured question, checked by a person on our team, or Google Analytics shows a visit coming from an AI. Nobody controls AI answers, so there is no guarantee of being recommended. The content states only hours and dishes you will serve, and the chef approves every piece. It promises no full dining room.

Questions

Won’t a text menu look worse than the photograph?

The photograph was never the answer. In the example, the assistants named the room that printed the allergen and the last seating, not the prettier PDF.

Do we have to list every allergen on every dish?

List the ones that are in that dish, in text. “Ask your server” is what the assistant cannot quote. A wrong list is worse than none: do not publish a guess.

Can a tourist list be the better answer?

Yes, when the question is only “where should I eat”. The small room wins the question that asks for a clock time and an ingredient.

How long until anything changes?

In the example, the first assistant named the room in week 6 and the second in week 10; two had not by week 12. Some questions never move.

Does a “best of” badge buy the answer?

It can buy a click. In the example it changed no assistant. The answer is written from a sentence, not from a badge.

See what AI answers about your business today.

One question, four AIs, free. The answers are saved as your starting point.

Run the free test