Why it matters

When the one choosing dinner is an assistant.

More and more people hand the finding, comparing and booking to an AI assistant. It doesn’t see the decor or hear the buzz; it checks facts. What gets a restaurant chosen now is clarity, not volume.

The leak

One restaurant, three versions online.

On one day in September 2026 we checked a real Korean restaurant in Allston, Boston. Google said 5:30 PM to 1 AM, every day. Its online-ordering page said closed on Friday and Saturday, with one item left on the menu. Two directories still linked a website that no longer exists. A person shrugs and calls. An AI assistant marks the place “cannot confirm” and moves on to the next one.

Three sources. Three answers.

Google Maps

Open daily 5:30 PM – 1 AM

Website: none

Online ordering page

Fri: Closed · Sat: Closed

Menu: 1 item

Restaurant directory

5:00 PM – 1:00 AM

Website: doesn’t load

One record, maintained

Fri 5:30 PM – 1 AM · Reservations by phone

Confirmed by the restaurant · date shown

Illustrative. Based on public listings we read on 2026-09-25. The restaurant is not a client and is not named.

How AI assistants choose

Six steps between “dinner for two near Allston” and a booked table. Restaurants fall out at four of them.

This is what an assistant actually does when someone hands it a dinner plan. We watched it happen.

01
Intent
“Friday 7:30, two people, one vegetarian, under $100, we’d like to book.”
No drop here
02
Discovery
Searches the web, maps and directories for candidates near the address.
Where restaurants get dropped
No website on your Google listing? You’re found through third-party pages that may be years out of date.
03
Reading
Opens each site or listing and pulls out hours, menu, prices, and how to book.
Where restaurants get dropped
Menu as a photo. Hours in an image. Prices missing. There is nothing to read.
04
Constraints
Checks every fact against the request: time, budget, diet.
Where restaurants get dropped
“Vegetarian?” Dish names don’t answer that. “Under $100?” No prices. You get marked “cannot confirm”.
05
Availability
Looks for a way to check the table: a booking link, a stated policy, or a phone number.
Where restaurants get dropped
No booking link, no policy. The assistant cannot finish, so it recommends the place that let it.
06
Action
Books, or hands the person a summary and the phone number to call.
No drop here
Observed in September 2026 by giving a real agent a real dinner task in Allston. It could not finish for any of the three candidates it found — not because they were bad restaurants, but because nothing confirmed price, dietary suitability or a way to book.

Assistants go where the data already is

When we gave a real agent a dinner task, it did not look for anything exotic. It searched the web, opened restaurant websites and directories, and used Google Maps to check hours, addresses and booking links. If your Google listing has no website, it learns about you from whoever wrote about you last.

No special file makes assistants find you. Being present in the sources they already use — your Google Business Profile, a working website, the booking platforms you actually use — is the whole of “discoverability”.

Restaurants are dropped for missing facts, not missing style

Once an assistant has candidates, it checks each fact against the request: time, budget, diet, party size, whether it can book. Every gap becomes a “cannot confirm”. In our test, all three candidates fell at this step. None was a bad restaurant. None had confirmed prices, dietary facts or a stated way to book.

This is the step we build for. Not to be ranked higher, but to not be dropped.

Listed is not confirmed

Hours on Google, a directory and an ordering page can all be different, and an assistant has no way to know which is right. A restaurant that states its own facts, with a date, resolves the tie. We keep “what listings say” and “what the restaurant confirmed” as separate fields, and we never quietly promote one to the other.

One record, four outputs

We maintain a single record of your facts. From it we generate the homepage people read, a facts page with sources and dates, a JSON data file with the same content, and a plain-text version. Change one field and all four change. Assistants can re-check the data file cheaply using standard web version tags, and we tested that they do.

Action is the last mile

Reading is not booking. We link to whatever you already use — phone, Toast, OpenTable, Resy — and state your policy in words. As booking and payment protocols for assistants mature, a maintained record is what they will connect to. We will add integrations when they are real, and say so when they are not.

How we fix it

One record. Every door.

You tell us a fact once. It appears on the page people read, on the facts page, in the data file assistants read and in the plain-text version — all at the same time, all carrying the date you confirmed it. Try it: change the Friday hours or a price.

Website · what people see

Marigold KitchenIndian · Somerville, MA

HoursFriday 5 PM – 10:30 PM

CallDirections

Confirmed by the restaurant on 2026-10-02

Facts page

Verification: Confirmed by the restaurant on 2026-10-02
Friday hours: 5 PM – 10:30 PM
Chana masala: $14 · vegan
Reservations: By phone · Call ahead for parties of 6 or more.
Source: Confirmed by the owner in writing. Fictional sample.

restaurant.json · what assistants read

{
  "schema_version": "1.0",
  "record_type": "restaurant_facts",
  "fictional_sample": true,
  "name": "Marigold Kitchen",
  "cuisine": "Indian",
  "merchant_verified_at": "2026-10-02",
  "hours": {
    "timezone": "America/New_York",
    "friday": {
      "opens": "17:00",
      "closes": "22:30",
      "closes_next_day": false
    },
    "kitchen_last_order": "22:00"
  },
  "menu": [
    {
      "id": "chana-masala",
      "name": "Chana masala",
      "description": "Chickpeas simmered with tomato, ginger and toasted spices.",
      "current_price_usd": 14,
      "vegetarian": true,
      "vegan": true,
      "contains": []
    },
    {
      "id": "butter-chicken",
      "name": "Butter chicken",
      "description": "Tandoori chicken in a mild tomato-cream sauce.",
      "current_price_usd": 18,
      "vegetarian": false,
      "vegan": false,
      "contains": [
        "dairy"
      ]
    },
    {
      "id": "garlic-naan",
      "name": "Garlic naan",
      "description": "Baked to order in the tandoor.",
      "current_price_usd": 4,
      "vegetarian": true,
      "vegan": false,
      "contains": [
        "gluten",
        "dairy"
      ]
    }
  ],
  "reservations": {
    "policy": "phone",
    "note": "Call ahead for parties of 6 or more.",
    "url": null
  },
  "ordering": {
    "online": false,
    "url": null
  },
  "address": {
    "street": "12 Elm Street",
    "locality": "Somerville",
    "region": "MA",
    "postal_code": "02143",
    "country": "US"
  },
  "phone": {
    "e164": "+16175550142",
    "display": "(617) 555-0142"
  },
  "data_updated_on": "2026-10-02",
  "source": {
    "kind": "owner_confirmation",
    "note": "Confirmed by the owner in writing. Fictional sample."
  }
}

facts.txt · plain text

Marigold Kitchen — restaurant facts (fictional sample)
Confirmed by the restaurant on 2026-10-02.

LISTED DETAILS
Name: Marigold Kitchen
Cuisine: Indian
Address: 12 Elm Street, Somerville, MA 02143, US
Phone: (617) 555-0142 (+16175550142)

FRIDAY HOURS
5 PM – 10:30 PM (America/New_York)
Kitchen last order: 10 PM

MENU
- Chana masala: $14; vegan
- Butter chicken: $18; contains meat
- Garlic naan: $4; vegetarian

BOOKING
Reservations: by phone. Call ahead for parties of 6 or more.
Online ordering: no

SOURCE
Confirmed by the owner in writing. Fictional sample.
Data updated: 2026-10-02

Two readers, one page

See what an assistant sees.

Assistants don’t see your photos or your typeface. They read structure: headings, lists, links, and the data behind them. We build pages that read well both ways.

Marigold KitchenIndian · Somerville, MA

HoursFriday 5 PM – 10:30 PM

CallDirections

Confirmed by the restaurant on 2026-10-02

Same fictional page, two readings. Nothing on the right was written twice — it is the page’s own structure.

What we tested

We tested with real AI agents before asking anyone to pay.

In September 2026, several agent clients — including Grok’s browsing tools and two script-driven bots — read our sample restaurant’s page and data file over the public web.
None of them invented a price, an opening time or a “vegetarian” label that had not been confirmed.
They correctly separated “listed publicly” from “confirmed by the restaurant”, and reported unknowns as unknown.
They re-checked the data file later using standard web version tags and detected a published change without being told about it.
Given a real dinner task, an agent could not complete it for any restaurant it found: every candidate lacked confirmed prices, dietary facts or a booking path.

Check my listing — free