Restaurant clients have started asking agencies a version of the same question: what do we do to show up when someone asks ChatGPT where to eat? It is a reasonable question, and it usually arrives with an assumption baked in, that somewhere there is a new channel to buy into, a file to upload, or a markup package that installs a restaurant into a generated answer.
None of those exist. What decides whether an assistant names a restaurant is the same local search record that decides whether it surfaces on a map, and the work involved is unglamorous enough that almost nobody is trying to sell it as a product.
Diners Are Already Asking, and Most Restaurants Are Missing
Research fielded by Dynata for DoorDash and SevenRooms in March 2026, covering 3,001 U.S. consumers, put the share of diners choosing restaurants with AI tools at 22%, while only 39% of operators had updated their listings to appear in those results. The behavior moved faster than the response to it, which is the gap an agency is being hired to close.
Restaurant discovery has never worked like ordinary SEO. It runs on proximity, the local pack, and business profile data rather than backlinks and page depth, and the gap between traditional and restaurant search is what explains where AI recommendations come from. The assistant is reading the same record everyone else reads. It just reads it faster and returns fewer names.
What the Assistant Is Actually Doing
In May 2026, Google published its first official documentation on visibility in generative AI features, and it is unusually specific about the machinery. Two mechanics matter for restaurants. The first is retrieval-augmented generation, which grounds answers in pages pulled from the Search index using core ranking systems. The second is query fan-out, in which the model issues a set of concurrent related queries to gather more than the original phrasing asked for. Google describes both in its guidance on grounding and query fan-out in Search, and the same page states plainly that a page must be indexed and eligible to appear with a snippet before it can show up in an AI feature at all.
Apply that to a realistic dinner prompt. Someone types a request for a place near the office that seats eight, has something a vegetarian will eat, and is open past nine. Fan-out turns that into separate retrievals for group seating, vegetarian options, the neighborhood, and hours. A restaurant that cannot be retrieved for any one of those components drops out of the synthesis before the model writes a word. Being generally well known in the market is not enough; the specific attributes have to be findable.
The Prompts Are Longer, and That Is the Whole Mechanism
Pew Research Center tracked 68,879 Google searches from 900 U.S. adults in March 2025 and found that query shape predicts whether a generated summary appears at all. Only 8% of one- and two-word searches produced an AI summary, compared with 53% of searches running ten words or longer, and 60% of queries that opened with a question word. Restaurant decisions are almost always phrased in that longer, constraint-heavy form, which puts them squarely in the band of queries that trigger a generated answer rather than a plain list of links.
The same study measured what happens next, and the finding reframes the goal. Pew recorded click behavior on results carrying AI summaries at 8% of visits, against 15% when no summary was present, and just 1% of visits produced a click on a source cited inside the summary itself. For a publisher, that is a traffic problem. For a restaurant, it is closer to a billboard: the value is in being named, because a diner who reads the name does not need to visit the website to act on it.
Most AI Optimization Advice Is Aimed at the Wrong Thing
Google used a section of the same documentation to name tactics site owners can ignore. It does not use llms.txt files. It does not require content broken into small chunks. It does not need copy rewritten in a special register for machines, and it treats manufactured mentions across the web as a poor use of effort. Most notably for anyone who has been sold a schema package, Google states that structured data is not required for generative AI search and that no special schema.org markup exists for it, while still recommending structured data for rich result eligibility in ordinary search.
What the documentation does point to for local businesses is Google Business Profile, which is the product that carries hours, categories, location, and menu links into AI responses. That should redirect the budget. A restaurant running an unverified or half-finished profile has a data problem no markup can compensate for, and the accuracy of the business details Google displays is doing more work in an AI answer than it ever did in a blue-link result, because there is no second and third listing for a diner to compare against.
Third-Party Platforms Are Answering on Your Behalf
The DoorDash and SevenRooms research also reported that 41.6% of restaurant-related AI queries cite third-party platforms as a source. Delivery marketplaces, review sites, and reservation aggregators are already well structured, heavily crawled, and consistently formatted, which makes them easy material for a model to ground on. When a restaurant maintains a thin first-party record, the assistant simply works from the aggregator copy instead, including whatever hours, menu, and photos the aggregator happens to be holding.
This is the familiar commission dynamic moved one layer up the funnel. A restaurant that has spent years watching aggregators outrank it for its own name now watches them supply the facts an AI repeats. The countermeasure has not changed much: make the first-party record more complete, more current, and more specific than the version a platform maintains, so the model has a better source to prefer.
Reviews Are the Trust Signal the Model Inherits
Models do not form independent opinions about food. They compress what the web already says, which puts reviews in a structural position rather than a reputational one. Michael Luca’s Harvard Business School working paper on one-star rating movements and restaurant revenue found that a one-star increase in Yelp rating produced a 5 to 9 percent revenue increase, that the effect was concentrated among independent restaurants rather than chains, and that a rating change carried roughly 50% more impact once a restaurant had at least 50 reviews.
That study predates generative search by more than a decade and measured human diners, but the input it identified is the one AI systems now aggregate on behalf of those diners. Volume gives a rating enough weight to be quoted with confidence. Recency and a visible response habit are what distinguish an active restaurant from a listing that may or may not still be open, and that distinction is exactly what a model has to resolve before it commits to a recommendation.
Constraint Questions Are Answered by the Menu
The prompts diners actually write are full of conditions. Gluten-free, vegetarian, a private room, a patio, a kids menu, open late, good for a group. Every one of those is answerable only if the corresponding fact exists somewhere as text a crawler can retrieve. A menu published as a PDF or a photograph of a menu board forfeits the entire category, along with every dish-level search underneath it, because the dish names and dietary language are not present as data.
Google’s position that no special markup is required for AI features is worth reading carefully here, because it is not an argument against structured data. Menu and Restaurant markup still earn rich result eligibility in ordinary search, and what structured data does for search visibility has not been diminished by AI features arriving alongside it. The more important point is that the underlying requirement is identical either way: real HTML text with dish names, descriptions, prices, and dietary terms written the way diners phrase them.
What to Measure, and What to Distrust
Search Console now includes a generative AI performance report covering how content surfaces in AI features on Search and Discover. Google pairs that release with a warning worth repeating to clients: no third-party tool has access to its internal ranking or AI systems, so vendor dashboards promising an AI visibility score are producing an estimate, not a metric. Treat them as directional and check their advice against the documentation.
The measurements that hold up on the restaurant side are the ones that were already worth tracking. Google Business Profile impressions, direction requests, and calls per location remain the closest available proxy for foot traffic. Review volume, average rating, and response rate show whether the trust layer is being maintained. Above those sits the harder question of durable visibility inside large language model outputs, which is a different standard from appearing once in a single answer and is the one worth benchmarking over quarters rather than weeks.
The Unglamorous Conclusion
There is no new line item here, which is a harder sell than a product but a more honest one. Restaurants that get named by AI are the ones with a verified and complete profile, a menu in readable text, business data that matches everywhere it appears, and a review flow that keeps moving. Agencies pitching restaurant clients on AI visibility are, in practice, pitching local SEO with a clearer reason to fund it. The work compounds, it is verifiable, and it does not expire the next time a model gets retrained.