AEO for restaurants: nobody asks for “a restaurant.”

A restaurant's hardest marketing problem used to be getting onto a list. The list was written by a critic or a guide or an aggregator, and once you were on it you were on it. Assistants do not work from lists. They compose an answer to a question that usually contains three constraints and no cuisine at all — fourteen people, Friday, somewhere with a set menu and a vegetarian main worth ordering. AEO for restaurants is the work of having published the facts those constraints are made of. And it starts with an unglamorous discovery: for most restaurants, the food itself is invisible.

The menu is a PDF, so the restaurant has no food

This is the defect that defines the vertical, and it has nothing to do with design taste. A language model composing an answer works from text it can read on the page. When the menu lives inside a PDF — or, more often than anyone admits, inside a photograph of a chalkboard or a designed JPEG — every dish, every price, every dietary notation, every “made in house” is outside the readable web. The restaurant exists as a name, an address, and a star rating. It has no food.

Now re-read the chart below with that in mind. Nearly every question in it is answerable only from menu detail, room detail, or policy detail. A restaurant whose menu is a PDF cannot be the answer to any of them, no matter how good the kitchen is or how many reviews it has.

The group booking is the most valuable answer in the vertical

Fourteen covers on a Friday with a set menu is a night, not a table, and the person asking is under mild social pressure to get it right. They will take a recommendation gratefully. What decides that answer is whether anyone has published the private dining details in text: how many the room seats, what the minimum is, whether there is a set menu and what is on it, how far ahead to book, whether the space can be closed off. Most restaurants keep this in an email template that a manager sends on request. In an assistant answer, an email template does not exist.

Dietary constraints are a hard filter, and precision wins them

“Gluten-free options available” and “we maintain a dedicated fryer” are two entirely different claims, and a diner with celiac disease is searching on the second one. The same holds for nut-free kitchens, halal preparation, and whether the vegan main is a real dish or a vegetable plate. This is a place where honesty is also the strategy: state what your kitchen actually does, including what shares equipment, in your own words. The cross-contact reality is the kitchen's to describe, not a marketing vendor's — and the precise version is both the safer one and the one that gets you into the answer.

Atmosphere is a query, and it is never on the menu

“Quiet enough to talk,” “nice but not a scene,” “where we can sit at the bar.” Diners ask for the feeling of the room constantly, and restaurant websites describe it in adjectives that every restaurant uses. The version that works is concrete: the room has twelve tables, the music sits under conversation, the back banquette is the quiet one, the bar seats eight and serves the full menu. Those are checkable facts, and checkable facts are what an engine can use.

What changed in how people choose

And 41% of consumers now always read reviews when browsing for a business, up from 29% the year before. The pattern is two-step: the assistant proposes two or three places with reasons, and the diner goes to read the reviews of those places before booking. Winning the proposal and losing the confirmation is a real way to lose the table.

What gets built, and in what order

• The menu as text, first and always. Every dish, every description, dietary notations, price ranges — on the page, readable. Keep the PDF alongside it for printing. Nothing else on this list matters as much.
• The private dining and large party page. Capacities, minimums, set menus, lead time, what can be closed off. The highest-value answer in the vertical and the most commonly missing page.
• The dietary page, in the kitchen's own words. What is prepared separately, what shares equipment, what can be accommodated with notice.
• The logistics facts in text. Hours including holidays, patio, dogs, parking, kids, noise, bar seating, whether the full menu is served at the bar.
• The entity pass. Name, address, phone and hours identical everywhere they appear, with restaurant and menu schema matching the visible page. Google states no special markup is required to appear in AI Overviews or AI Mode — schema is here for unambiguous parsing and classic rich results, and the menu text still has to be on the page.

Each arrives as a draft in an approve queue — one prepared move per day, nothing published until someone at the restaurant approves it. Menus change; the queue is how the site keeps up without becoming somebody's second job.

How you know whether it worked

Scans re-ask the same diner questions on a schedule, so the record is a before and after. Approved pages are submitted to Bing within minutes of going live, and Bing standing is the fastest indicator — independent testing found roughly 87% of ChatGPT's citations match Bing's top results (Seer Interactive, 2026), a measured correlation and not a mechanism OpenAI documents. Citations run slower: a median of about 6.8 days after indexing, roughly 42% of pages within 30 days, in practitioner testing across 2026 (Semrush / practitioner testing, 2026). No date is ever promised. And 40–60% of the sources AI cites change every month (eMarketer, 2026), which in a vertical with seasonal menus is an argument for keeping the text current rather than a reason to despair.