Search engines now find open tables and order food for guests. The booking rails exist. What they can't fix is a restaurant whose hours, menu and details disagree from one listing to the next, and whose sold-out dish is known only to the POS.
“Find an Italian restaurant near me with a table for four at 7:30 tonight, gluten-free options and parking.” A guest says it in one breath. An agent has to resolve six facts: the cuisine and location, tonight's hours, the menu, the dietary statements, an open table and the parking. In most restaurants those come from different systems with different owners.
Google's AI Mode finds reservations with live availability across reservation platforms, and since November 2025 any U.S. user can book through it. In August 2026 Google Maps added food ordering to Ask Maps, with partners like Square and Toast. The agent can reach the reservation book and the ordering system.
That's the action. The question before the action is still whether the restaurant it found is the right one.
A location's hours live on the site, the Google profile and every ordering and reservation listing. A new menu goes into the POS and reaches the PDF, the Google menu and the delivery apps when someone remembers; Google itself says menu edits take a day or two to show. Some Google attributes can come from customers, not the owner.
Multiply that by locations. A group can be right at one door and wrong at the next, and the agent has no way to tell which listing to trust.
In October we audited Guadalupe Mountain Brewing in Carlsbad, New Mexico, with the owners' approval. Eight sources described it. Their hours came out four different ways. Two street addresses and two phone numbers were in circulation, and the name was spelled three ways. One brewery directory listed another brewery at its address. The website returned an error on every page. Asked about the best brewery in Carlsbad, an AI assistant answered for Carlsbad, California.
None of that is a menu problem. It's the location truth underneath the menu, and it's exactly what a guest's agent reads first.
When the kitchen 86es a dish, the POS knows. Square lets a seller mark an item sold out at one location with a time it comes back. DoorDash treats 86ing as an immediate menu update. But each system holds its own copy, and a guest's agent reading the menu page tonight sees none of them.
The rule we built: an item's state tonight comes only from a current reading, under 15 minutes old. Older is “ask the restaurant.” Missing is unknown, never “sold out.”
A careful agent repeats what the restaurant says about a dish, in its words, with its kitchen caveat: “gluten-free, prepared in a kitchen that uses wheat.” It never infers “free of” from silence, and it never calls a dish safe for an allergy. Allergy questions end with a person on staff.
The data gaps make that rule necessary. Schema.org has no allergen field, and the major allergen list grew to nine with sesame in 2023.
Booking a table and ordering food run through the restaurant's own systems. Lehvel's first action is smaller: a request about a large party, private dining or catering, routed to a person at the right location, with a receipt that commits nothing.
Booking, waitlists, deposits and payments through an agent via Lehvel each need the restaurant's opt-in, its own system's interface and a gate in front of them. Today the agent is sent to the restaurant's reservation page.
Restaurants are the clearest case of a pattern we see in every business customers visit or book. Business Truth says which location and when. Menu truth says what's served. Operational truth says what's available now. The booking platform does the booking.
Get the first three right once, with a source and a time on every fact, and every assistant that asks gets the same answer.