Getting your restaurant into AI answers when someone asks for recommendations

The pitch you are getting in your inbox says otherwise. It says there is a new channel, a new file to publish, new markup you are missing, and that you are already behind. Most of that is invented.

What is actually true is smaller and more useful. The kind of question people now ask out loud, "where should I take eight people for dinner in Alameda on a Thursday," is a question your listing either answers or does not.

How does an AI actually pick which restaurants to name?

It retrieves, then it summarizes.

Google's own documentation describes its AI features as built on core Search ranking, pulling from the same index, using retrieval and grounding plus what Google calls query fan-out, "a set of concurrent, related queries generated by the model." A page has to be indexed and eligible to be shown with a snippet. Google states there are no additional technical requirements beyond that.

For local questions, Google says specifically that a Business Profile helps a business appear "in both AI responses and other Google Search results." Same record, two surfaces.

So the mechanics are not new. What is new is the shape of the answer. Instead of ten listings, the customer gets three names and a sentence about each. That sentence has to come from somewhere, and it comes from whatever the internet already says about you plainly enough to be repeated.

If nothing on the internet describes your restaurant in a specific sentence, nothing will be summarized into one.

Whitespark's testing of AI Mode local answers in May 2026 found that star ratings, review volumes, pricing and hours are pulled from Google Business Profiles or comparable sources, and that reviews and unstructured citations, meaning blogs, news and community pages, are the heaviest inputs. That is practitioner observation from a vendor, not a controlled test. It points at the same two assets anyway: the profile, and what other people have written about you.

How much of your local demand is actually an AI answer?

Less than the panic suggests, and the arithmetic is worth doing before you spend a dollar.

Ahrefs analyzed 146,122,391 desktop SERPs from September 2025 and found AI Overviews on 20.5% of all searches, but only 7.9% of local searches. Sterling Sky, tracking separately with different tooling, observes AI-powered local packs on roughly 7% of tracked keywords. Two methods, two vendors, the same neighborhood.

Here is what that means in slots. Take 1,000 local searches a month where your restaurant could plausibly surface.

• About 79 of them return an AI answer. The other 921 return the map pack you already understand.
• Those 921 searches offer three local pack slots each: 2,763 slots.
• On the AI side, Sterling Sky reports AI local packs show one or two businesses instead of three. Call it 1.5. That is about 119 slots, where the old 3-pack would have given 237.
• Net change: roughly 118 slots lost out of 3,000, or about four in a hundred.

That is a real contraction and it is not an emergency. Places Scout's analysis behind Sterling Sky's report counted 5,943 unique businesses in AI local packs against 18,330 in regular 3-packs, about 32% as many, and found that in 322 markets, 88% had fewer unique businesses in the AI version. Fewer restaurants get named. That is the actual risk, and it is a slow one.

Now the honest counterweight. A Whitespark study of 540 manual queries across six local industries reportedly found AI Overviews on about 68% of local business queries, and 97% for blended questions like "best place for a birthday dinner downtown." Both findings are right about their own sample. Ahrefs classified local by intent across a database dominated by short "[cuisine] [city]" patterns. Whitespark hand-picked long conversational questions.

The synthesis: short transactional searches are still local pack territory, and long conversational "help me choose" questions are AI territory. Your customers do both, and the second kind is growing.

Is there something special I have to publish?

No, and this is the part worth being blunt about.

You do not need a file at the root of your site. Google explicitly names llms.txt and special markup among the things that do not work. You do not need structured data either. Google states that structured data is not required for its generative features and there is no special schema for them.

You also do not need to rewrite your pages into artificial chunks, and you should not go buying mentions on directories nobody reads. Google names inauthentic mentions and content rewritten for AI systems as ineffective.

Google's own sentence is the one to keep: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

Does an llms.txt file do anything?

No, and this one has been measured rather than argued.

Ahrefs examined every domain in its Web Analytics product with traffic in May 2026, 137,210 sites, verified which served a real llms.txt file, then looked at every request to llms.txt paths across the whole population by user agent and response code.

• 28% of those sites publish an llms.txt.
• 97% of those files received zero requests in May 2026, from any bot or any human.
• Of the traffic that did occur, 96% was bots and 4% humans, and 77% of those bots were not AI tools at all.
• AI retrieval bots, the ones that actually produce citations, were 1.1% of requests.
• Slackbot fetched llms.txt files more often than PerplexityBot did.
• Zero AI bots requested files that did not exist. Of requests returning a 404, 98% were human. Publishing one puts you on no list. Not publishing removes you from none.

Ahrefs sells SEO software, and its own caveat is worth repeating: fetched does not mean read, so every figure is a ceiling.

Google's mythbusting page agrees in its own words: you do not need to create machine readable files, AI text files, markup or Markdown to appear in Google Search, and doing so will neither harm nor help, because Google Search ignores them.

If a proposal includes an llms.txt line item, that line item is worth nothing. The rest of the proposal may still be fine. Ask what else is in it.

The category sells a new discipline because a new discipline is easier to bill than an accurate hours field.

Why does it name my competitor instead of me?

Because their facts are easier to use.

A summarizer answering "good spot for a birthday dinner with a group" wants a restaurant that has published something about groups. Large party seating, a private room, a set menu, a headcount. If your competitor published that and you did not, they get named, regardless of which room is nicer.

Three things separate the restaurants that get named.

Specificity. Not "great food and warm atmosphere." A patio with heaters, a room that holds twenty, a kitchen that closes at eleven, a gluten free fryer. Concrete, checkable facts.

Consistency. Your hours, address, name, and cuisine should say the same thing on your site, your profile, your reservation platform, and every marketplace that lists you. When sources disagree, a summarizer either picks wrong or skips you.

Recency. Current photos, current hours, and reviews still arriving read as a business that is currently operating. A stale record is a risky thing to recommend.

Why does the answer change every time I check?

Because these systems are not deterministic, and the variance is larger than most people believe.

Researchers at the University of St. Gallen ran daily prompts across ChatGPT, Gemini, Google AI Mode and Perplexity over a 45 day window in early 2026, plus repeated same-day runs of identical prompts. Across 4,044 consecutive-day pairs, roughly 65% of the cited sources changed overnight. Running the same prompt ten times on the same day produced source overlap of 32% to 43%. ChatGPT was the least stable of the four.

Their prescription is specific and you should hold any vendor to it: at least seven runs per prompt per day, reported on a two to four week rolling aggregate, never week over week, across a varied prompt set.

What that means for you as an operator: checking once proves nothing. If you asked on Tuesday and got named, and asked on Wednesday and did not, nothing happened to your restaurant. An agency reporting that your AI visibility score rose five points this week is reporting noise.

What actually makes a restaurant easy to recommend?

Answer the questions people ask out loud, on your own site, in their words.

Write the page a person would need if they asked by phone. Group dining. Dietary restrictions. Parking. Late kitchen. A party of twelve on short notice. Plain pages, plain sentences, real answers.

The one piece of published research on generative optimization supports a narrow version of this. The KDD 2024 paper that introduced the term found that adding verifiable statistics, direct quotations from credible sources, and inline citations produced the largest gains. It also found keyword stuffing actively hurt, scoring below the do-nothing baseline, and that an authoritative tone produced no significant improvement. Note the limits before anyone sells a package on it: the experiment ran over five fixed sources with an older model, and nothing in its benchmark was a local commercial query.

Keep the menu as readable text, with dish names and prices, and keep it current when the menu turns. A menu that goes stale each season is a recurring source of wrong answers, which is why handling seasonal menu changes on the website is worth building a routine around rather than doing in a panic.

And remove the things that make your site hard to read at all. A homepage that is one video, a menu locked in a PDF, and a contact form instead of a phone number will keep you out of answers you should be winning, which is a subset of the website mistakes that cost covers.

Demand also moves in cycles, and the questions change with them. Graduation, holidays, and corporate season each bring a different kind of ask, and being the specific answer during those windows is worth more than being a general answer year round. That pattern is mapped in the seasonal cycles behind catering demand.

What to do this week

Write down the sentence you would want an AI to say about your restaurant. One sentence, the way you would tell a friend.

Then find every fact inside that sentence and check whether it is published anywhere a machine can read. Not implied by your photos. Written in words.

Then fix the disagreements. Hours, name, address, cuisine, price band, service options, across your site and every platform that lists you.

Then publish two pages that answer the two questions you get most on the phone, in the words customers use, not in the words your menu uses.

Then leave it alone for a few months. Chasing this weekly wastes an owner's attention, and the numbers that tell you whether marketing is working are slower and duller, which is the point of the marketing metrics that predict cover count.

Questions about how AI describes your restaurant?

Email me at eric@seod.com with one sentence: the thing you would most want an AI to say about your restaurant when someone asks for a recommendation. I will tell you which facts in that sentence already exist publicly, which ones do not, and where I would publish the missing ones. Usually a short list of two or three gaps.

I ran restaurants for sixteen years and I answer these myself. If the honest read is that your profile already supports the sentence and you just need patience, that is what you will get back.

Or keep reading more on restaurant marketing.