Agent Policy for Restaurants and Food Delivery
Published November 14, 2026 · Reviewed November 2026 · By the Ingenious AI team
The short answer
A restaurant's agent traffic wants two things: answers (menus, hours, dietary information, locations) and actions (reservations, orders). The restaurant pattern welcomes the first completely and fences the second exactly like a store fences checkout: book and any order-placing are refused or held at conditional, while menu reading is the freest grant in any vertical — a menu an agent cannot read is a restaurant its users never hear about.
What is the restaurant split?
Information generosity, transaction caution, at small-business scale. Menus are perishable public documents — they change with seasons, markets, and the chef's morning — and their whole purpose is to be consulted before a decision. Agent consultation is consultation: a user asking "which places near me have a tasting menu under this price, with a vegan option?" is reading menus at one remove. Ordering and reserving, by contrast, commit a kitchen and a table; errors land on a Friday night, in person. The policy states both halves in a file short enough for an owner-operator to maintain between services.
Menus, hours, and the freshness habit
Grant browse, read, search, and quote on the catch-all at reading pace, and put the freshness sentence in the data-use terms: menus and prices as published today, specials confirmed with the restaurant. Restaurants change offerings faster than any cache rhythm the other verticals use, so the terms matter more than the cap: an agent quoting last month's menu confidently is worse than one that could not read it at all. The operational corollary is yours, not the agent's: the published menu must be the current menu, which makes the policy review a by-product of the menu update — when the menu changes, the file's last_updated is already in your hand.
Reservations and orders: the fence
book is the reservation verb and belongs refused or conditional in every rule a restaurant writes for the open world: a phantom reservation costs a real table, and no-shows are already the industry's tax without automated help. Delivery and takeaway ordering map to shop plus the committing step — treat the commit as checkout and fence it identically: agents may assemble an order with their user (the menu work is the valuable part), and the order itself is placed by the person or through the restaurant's own ordering channel. Named delivery platforms are the exception layer — their rules reflect the commercial agreements the restaurant already has with them, recorded, as always, not created, by the file.
Small site, real policy
Restaurants are the vertical where owners most often assume the format is for larger businesses. The arithmetic argues otherwise: the restaurant pattern is 2 rules and a notice — an evening's work with the free generator — and the audiences are already arriving, because "where should we eat?" is among the most agent-asked questions in consumer life. A single-location restaurant's file will be the shortest in this guide series and may do the most direct commercial work per line: discoverability in agent answers, on terms the owner wrote, with the table protected.
A last, practical note for owner-operators: your policy interacts with the platforms your customers already use, whether or not you ever write a file. Assistants answering "where's good near here?" work from whatever they can read — platform listings, review sites, your menu if you publish it. The policy does not create that reality; it gives you authorship inside it: your menu as the source, your freshness terms around it, your booking channel as the stated route. Restaurants that treat the file as one more listing to keep current — alongside the menu board and the platform profiles — find it unremarkable within a month: same information, one more place it lives, with the distinct advantage that this place carries your terms.
Worked example: a bistro group's menu-first policy
Fictional example. Mesa Verde Hospitality, a fictional group running 6 bistros, publishes 1 file per location domain, each with 3 rules: the catch-all granting menu reading free at 30 requests per minute with the freshness sentence; a named rule for its reservations partner permitting book under the existing agreement, cancellations excluded from agent actions entirely; and a delivery-platform rule mirroring its takeaway contract. book is disallowed in the catch-all, and ordering commits are refused in every rule but the platforms'. The group's measured first year: agent-referred discovery became a tracked booking source at 2 locations; phantom reservations attributable to agents, 0; menu staleness complaints, 3 — each traced to a location that had updated its printed menu a week before its website, a discipline the policy review now catches.
The restaurant checklist
- Free the menu. Reading and quoting granted on the catch-all — discoverability is the point of publishing a menu.
- Write the freshness sentence. Menus as published today; specials and prices confirmed with the restaurant.
- Fence the table.
bookrefused or conditional for the open world; partner reservations as named rules under real agreements. - Treat ordering as checkout. Agents may assemble; people (or your own channel) commit.
- Review with the menu. The file's date moves when the menu does — the cheapest review rhythm in any vertical.
Keep reading
Agent Policy for Schools and Universities · Agent Policy for Marketplaces · Agent Policy vs robots.txt, Compared · Spec samples · Conformance · Pricing · All guides
Frequently asked questions
We only take phone reservations. What should the policy say?
That, plainly: book disallowed in every rule, and the notice carrying the sentence agents can relay — reservations by phone, with the number. An agent that reads your file and tells its user "they take bookings by phone, here is the number" has done your customers a service your old silence never could.
Should delivery platforms have special rules?
If you have commercial agreements with them, yes — a named rule per platform reflecting the agreement's agent-facing shape: what may be ordered, at what pace, under what data terms. The policy records the arrangement; the contract remains the arrangement. Platforms you have no agreement with live under the catch-all like everyone else.
Our menu changes daily. Is a policy pointless?
The opposite: daily change is why the freshness sentence exists. The policy does not freeze your menu; it tells agents the menu they read is today's, must be re-read rather than cached, and yields to the restaurant's word. Agents built for declared terms plan accordingly — short cache, confirm specials — which is better behaviour than any scraper ever offered you.
Can agents answer dietary and allergen questions from our site?
They will answer from what you publish, so publish with that reader in mind: clear allergen information on the menu pages, and data-use terms requiring that dietary answers quote it rather than improvise. The policy cannot certify accuracy — that is your kitchen's and your menu's work — but it can insist the agent's source is your text, attributed and current.
Do we need a policy per location?
Per domain, yes: each origin publishes its own file. Groups running several locations on one domain publish one file whose rules cover the group, while separate domains each carry their own — usually from one template with per-location contacts and partners, reviewed as a small fleet.
Generate a policy — free · Validate a policy · What an agent access policy is