What guests really ask the AI: an analysis of typical restaurant prompts

Why the restaurant search moved into the chat box

The query got longer and the answer got shorter. What used to be "italian restaurant downtown" typed into Google, followed by ten links to sift through, is now a whole sentence typed into ChatGPT: "Table for four on Saturday around seven, somewhere Italian and not too loud, one of us is vegetarian." What comes back is two or three names, not a ranked list. There is no page two to be found on, and the click that used to follow often never happens at all. In Pew Research Center's browsing data from 900 US adults, people clicked through to a website in 8% of searches where an AI summary appeared, against 15% where none did, and clicked a link inside the summary itself in 1%. Google disputes that methodology, but the direction matches what SparkToro measured across US Google searches in early 2026: 68% ended with no click to any site. The answer is the storefront now. You are named in it, or you are invisible to that guest.

Eating out suits this format unusually well. Time of day, occasion, party size, one person's diet, roughly what you want to spend: five conditions a guest can state in one breath and have filtered in one pass, instead of opening eight tabs and comparing. That is also where your name is won or lost. None of the engine makers publish the rule that decides which restaurants get named, so the work that pays sits on your side of it. Know which questions are being asked, and make sure the facts each question needs are findable.

Where do those names come from? For questions like these the major engines retrieve before they write. They pull from Google Business Profiles, review sites, menu pages, reservation platforms, local directories and your own website, then summarize what they found. They are not careful about it either: when Columbia Journalism Review's Tow Center put 1,600 source-identification queries to eight AI search tools, more than 60% of the answers were wrong, and ChatGPT expressed any uncertainty in only 15 of 200 attempts. So appearing in an answer is not luck, and being described correctly is not automatic. Generative Engine Optimization, GEO for short, is the unglamorous work of fixing what those sources say about you.

The three prompt patterns behind most restaurant questions

Read enough of these questions and the shapes repeat. The most common is a place plus an occasion: "where can we get a nice dinner for two downtown?" or "somewhere near me that can take ten people for a birthday." Location, party size, mood. Your name only comes up if all three are findable: which town or neighborhood you are in, that you can seat a table that size, and some sign, in your words or your guests', that a big group is welcome rather than merely tolerated.

The second pattern is diet. "Which restaurants here do a proper vegan main?", "where do I get gluten-free pizza?", "anywhere doing lactose-free?" This is what the format is genuinely better at than a page of links, because filtering on one detail buried in a menu is exactly what a language model can do, provided the detail is written down somewhere it can read. A kitchen that cooks vegan happily but never types the word cannot be recommended for it. The engine is matching words, not tasting food.

The third is logistics: open on a Sunday, walk-ins without a reservation, a terrace, a high chair, step-free access, parking. Yes-or-no facts, either recorded or absent, and absent usually means left out rather than guessed. A model assembling a shortlist of places open on a public holiday has no reason to gamble on you when nothing in its sources says you are open.

What an answer engine can actually read about you

A language model does not browse your site the way a hungry person does, and it has no opinion about your photo of the burrata. What it can use are facts in text: cuisine, price band, hours, address, whether you take reservations, what makes the place particular. Put those in a picture of the chalkboard or a PDF menu and, for practical purposes, they are not there. Marking up hours, address and prices with schema helps as well. Not as a ranking trick, since Google has said since 2018 that structured data is not a direct ranking factor, but because a tagged fact is harder to garble than a sentence. Readable text beats a beautiful layout. That is the unromantic core of GEO for restaurants.

The most underrated asset is your guests' own vocabulary. People ask for "quiet", "romantic", "good with kids" or "proper home cooking", soft words that can only match if the same soft words exist in text about you, and reviews are usually where they live. There is real evidence behind this: across roughly 75,000 brands, Ahrefs found that how often a brand is mentioned around the web correlated with how often AI tools cite it at about 0.66, some three times the correlation for backlinks. No engine publishes its actual selection rule, so treat that as a strong hint rather than a formula. Either way, what other people write about you is a bigger share of your AI visibility.

Then there is consistency, the dull half of the job. If you are "Ristorante Bella" on Google, "Trattoria Bella" on Facebook and "Bella Bar & Kitchen" on your own site, and those three disagree about Sunday hours, the model has to pick one version or hedge. Neither outcome helps you, and the guest who arrives at a locked door blames you, not the software. Nobody has published a penalty for contradicting yourself. But every source that agrees with the others is one fewer way for the answer to come out wrong.

Three real prompts and the facts each one needs

Take "I'm here for work and want somewhere quiet for a lunch meeting." Landing in that answer needs three things on the record: that you serve lunch on a weekday, that the room is calm rather than lively, and that a table of two with laptops is ordinary here. A homepage that says "we look forward to welcoming you" supplies none of them. "Quiet tables at lunch, fast WiFi, used to business lunches" supplies all three in ten words, and reads no worse to a human being.

Or the plain one: "where can I get a good schnitzel near here?" It sounds too simple to teach anything, and it is the clearest case of all. The dish has to be named in text. If your menu page carries the words "Wiener schnitzel, veal", you are a candidate. If your menu is a photo, or lives only inside a delivery app, you are not, however good the schnitzel is. None of this rewards cooking. It rewards having written the thing down.

Then the compound one: "somewhere with vegan mains and outdoor seating for Friday night." Three conditions that all have to hold at once. These are the easiest answers to win, because most of your competition fails on exactly one missing attribute: the terrace nobody wrote down, the vegan main that only exists on the specials board. Every attribute you record properly adds another combination you can be the answer to.

Reviews: the words other people use about you

Reviews do two jobs at once. They are a verdict, and they are a body of text about you written in the words guests actually use. If yours are full of "huge portions", "fair prices" or "the room is lovely", those phrases are there to be matched against questions phrased the same way, including qualities you would never think to claim in your own copy. The job is not to manufacture them. It is to ask happy guests at the table, while it is fresh, and to ask for something concrete: what they ate, what they were celebrating.

Your replies are public text too. Answering a complaint plainly and thanking someone for praise builds a record that reads like a restaurant paying attention; snapping back at every two-star review builds a different one, and both are equally readable. So ask for details rather than stars. "Mention the lamb and the terrace if you liked them" costs a guest nothing, and it puts the exact words a future prompt will contain into a source an engine can reach.

Your menu is the most quotable page you own

Your menu page is the most fact-dense thing you have: dozens of dish names, ingredients, prices and dietary labels on one URL. Plenty of restaurants then bury it in a PDF, a photo, or a third-party ordering widget that renders no readable text at all. A person will squint and cope. A crawler gets nothing. Put the menu on your own site as real text, with dish names, a line of description, and honest labels for vegetarian, vegan, gluten-free and spicy. Every label is a question you become answerable for.

Write dishes the way guests describe them. "Pasta 4" answers nothing. "Homemade tagliatelle with porcini, vegetarian" answers pasta, homemade, vegetarian and mushrooms in a single line. Say where things come from and when they are on, because "local" and "seasonal" get asked about constantly. One caveat worth respecting: Google's own guidance warns against writing separate content "for AI" and treats mass-produced pages as spam. This is not that. This is describing your food accurately, which happens to be the version a machine can read too.

Then keep it current. A dish you dropped in spring, or a price from two years ago, will be repeated back to guests as fact, and price and opening hours are exactly the kind of detail AI answers garble. Clean structured data on your own page is the cheapest way to shorten those odds. Stale menus produce disappointed arrivals, disappointed arrivals produce worse reviews, and worse reviews feed straight back into what gets said about you. If the kitchen changes with the season, the page changes with it.

Your Business Profile, and the facts that go stale

After your own site, your Google Business Profile is the densest sheet of facts about you that anyone can read, and the one most likely to be quoted back. Hours, cuisine category, price band, terrace, step-free access, reservations, takeout, service options: fill in every field the profile offers, including the ones that feel trivial. Half-empty profiles are the norm, which is why so many restaurants never surface for specific questions. Each attribute you complete is one more question you can answer instead of a gap somebody else fills.

Two things deserve a calendar reminder: holiday hours, and current photos. Anyone asking what is open today on a public holiday gets answered from whatever hours are on file, so the profile you left alone in March decides whether you appear in December. Then run the same sweep across review sites, reservation platforms and local directories, so the name, address and phone number match everywhere. The trade calls it NAP consistency. No engine has confirmed it as a ranking signal, but it removes contradictions, and contradictions are what make a model hedge, or repeat the wrong version of you.

Where to start, and how to check your work

Start with three things, in this order. One: put the full menu on your site as real text, with dietary labels. Two: finish the Google Business Profile, every attribute plus the holiday hours. Three: spend the next few weeks asking happy guests for specific reviews. None of it needs an agency, a budget or a new website, and all three fix the same underlying problem, which is facts about you that no machine can currently read.

After that, audit yourself. It takes an evening. Open ChatGPT, Gemini and Perplexity yourself and type the questions your guests would type: your cuisine, your neighborhood, your occasions, the awkward ones about diets and Sunday hours. Are you named? Is anything about you wrong? A wrong answer is more useful than a missing one, because it tells you which source to go and fix. Repeat it a couple of times a year, and know that not every answer is even looked up: in Profound's analysis of roughly 730,000 US ChatGPT conversations, only about 18% triggered a web search at all. The rest came from what the model already held, which moves on its own schedule.

GEO is maintenance, not a project. The restaurant whose hours, menu and labels are accurate this week, written in the words its guests use, gets described correctly more often than the one relying on the food to speak for itself. That is a smaller claim than the hype makes, and it is the one that holds up: as more tables begin with a typed sentence, being legible is the difference between making the shortlist and never being mentioned.