Can an AI agent book or buy on a local business website?
AI assistants are starting to book tables, appointments and rooms for people. We read the 694 websites of local businesses that AI assistants recommend, in 20 countries and 22 kinds of business, the way a simple agent would: without running scripts, logging in or submitting anything.
Published by Guglielmo Paesano, founder of AIGROW · working paper, 5 November 2026
- Countries
- 20
- Kinds of business
- 22
- Sites it could read
- 571
- Nothing an agent can act on
- 47.4%
Working paper, not peer reviewed
Can an AI agent book or buy on a local business website? AIGROW Labs read 694 websites of local businesses AI assistants recommend, in 20 countries, as an agent would. 27.1% let it book or buy through a standard system: 52.8% of restaurants, 9.5% of other trades. 47.4% gave it nothing to act on.
Findings
Each share is of the 694 sites unless it says otherwise. The tables below give every figure with its 95% interval.
About one site in four lets an AI agent book or buy through a standard system: 27.1%, 188 of 694.
A standard system is a known booking, ordering or shop service embedded in the pages or linked from them, shop software with a working cart, or a declared booking action. 52.8% of restaurants reach it, 38.6% of hotels and 9.5% of the other trades. The share is a lower bound: a reader that runs no scripts misses booking widgets drawn after the page loads, and counting every own booking form, cart signal and empty booking frame raises it to 35.4%.
Nearly half give an agent nothing to act on: 47.4% of the sites.
17.7% could not be read at all, most often because the server refused the reader. 29.7% could be read but had nothing to fill in and no system to use, at best a phone number or an email address to hand back. Outside restaurants and hotels the share with nothing to act on is 57.8%.
Only 3.7% of sites close robots.txt to an AI agent by name, but 8.9% refused our openly named reader at the door.
Every site that closes its home page to an AI agent by name closes it to a crawler that collects training data, and 0.6% close it to the agents that fetch a page because a person asked. The refusals come from the server or its security service, and they shift with the reader: of 556 sites that served a browser-like reader the same day, 27 turned ours away, and of 60 that refused it, 21 let ours in.
What it means for a business
An assistant that cannot get into your site, or finds only a phone number, hands the person back a name and nothing more. Three things decide whether it can go further.
- Let it in
- Check what your server sends an AI agent, not only what robots.txt says: here the firewall turned away more readers than robots.txt did, and a block there goes unnoticed because the site still opens in a browser. The free scan checks both for the assistants’ crawlers, and the free crawler check does it for any page.
- Give it a way to act
- If you take bookings or orders through a service, link to it from your home page. A reader that runs no scripts finds a plain link and can miss a widget drawn after the page loads.
- Write the facts down
- Say in your page markup what kind of business you are and when you are open: of the readable sites, 25.7% described themselves as a local business and 13.0% gave their opening hours. AI Local Fix adds LocalBusiness markup to up to 10 of your pages as part of its groundwork.
Where the sites stand
Each site sits on the highest step an agent reached on any page read. Bars are shares of all the sites in the row, so a full bar would be every site.
- Book or buy through a standard system
- Nothing an agent can act on
The ladder by sector
Share of each sector’s sites on each step, with its 95% interval.
| Sector | Sites | Cannot read | Call or email only | Labelled form | Standard system |
|---|---|---|---|---|---|
| All sites | 694 | 17.7% (15.1% to 20.7%) | 29.7% (26.4% to 33.2%) | 25.5% (22.4% to 28.9%) | 27.1% (23.9% to 30.5%) |
| Restaurants | 197 | 17.3% (12.6% to 23.1%) | 20.3% (15.3% to 26.5%) | 9.6% (6.3% to 14.6%) | 52.8% (45.8% to 59.6%) |
| Hotels | 127 | 22.0% (15.7% to 30.0%) | 10.2% (6.1% to 16.7%) | 29.1% (21.9% to 37.6%) | 38.6% (30.6% to 47.3%) |
| Other trades | 370 | 16.5% (13.1% to 20.6%) | 41.4% (36.4% to 46.4%) | 32.7% (28.1% to 37.6%) | 9.5% (6.9% to 12.9%) |
On the top step, 89.4% use a service from the list, 61.2% embed it in the page and 28.2% only link to it; shop software with a working cart accounts for 19.1%.
Why 123 sites could not be read
The largest reason is a refusal: an error status such as 403 or 429, or a bot challenge page. Shares are of all 694 sites.
| Reason | Sites | Share of all sites |
|---|---|---|
| Refused, or answered with a bot challenge | 62 | 8.9% (7.0% to 11.3%) |
| No readable text without scripts | 20 | 2.9% |
| The name does not resolve | 14 | 2.0% |
| Error status on the home page | 8 | 1.2% |
| No answer within 15 seconds | 6 | 0.9% |
| robots.txt answers with a server error | 4 | 0.6% |
| robots.txt closes the home page | 3 | 0.4% |
| The secure connection fails | 3 | 0.4% |
| The home page is not a web page | 1 | 0.1% |
| Parked domain | 1 | 0.1% |
| Server placeholder page | 1 | 0.1% |
robots.txt and the firewall
Shares of the 668 sites whose robots.txt answered. Few owners close the door to AI agents on purpose, and almost none to the agents that act for a person.
| robots.txt | Sites | Share |
|---|---|---|
| robots.txt names at least one AI agent | 70 | 10.5% (8.4% to 13.0%) |
| Closes the home page to an AI agent by name | 25 | 3.7% (2.5% to 5.5%) |
| to a crawler that collects training data | 25 | 3.7% (2.5% to 5.5%) |
| to a crawler for AI search | 2 | 0.3% (0.1% to 1.1%) |
| to an agent that fetches a page for a person, counting the rules for every robot | 4 | 0.6% (0.2% to 1.5%) |
| The rules for every robot close the home page | 3 | 0.4% (0.2% to 1.3%) |
At the firewall, 8.9% of all 694 sites refused the reader with an error status or a challenge page. Of the 556 sites that served a browser-like reader the same day, 4.9% turned ours away; of the 60 that refused it, 35.0% let ours in.
What a readable site tells an agent
Shares of the 571 sites that could be read. Most carry some markup, but few state the facts an agent needs to book.
| On the pages read | Sites | Share |
|---|---|---|
| A tappable phone link | 357 | 62.5% (58.5% to 66.4%) |
| An email link | 328 | 57.4% (53.4% to 61.4%) |
| A request form whose fields all carry a readable name | 259 | 45.4% (41.3% to 49.5%) |
| Only forms whose fields carry no readable name | 58 | 10.2% (7.9% to 12.9%) |
| Its own booking form, with date and guest fields | 57 | 10.0% (7.8% to 12.7%) |
| An empty booking frame left for a script to fill | 28 | 4.9% (3.4% to 7.0%) |
| Any schema.org markup | 450 | 78.8% (75.3% to 82.0%) |
| Describes itself as a local business in markup | 147 | 25.7% (22.3% to 29.5%) |
| Opening hours in markup | 74 | 13.0% (10.5% to 16.0%) |
| Says in markup whether it takes reservations | 13 | 2.3% (1.3% to 3.9%) |
| Declares an action to reserve, order, buy or schedule, with a target | 13 | 2.3% (1.3% to 3.9%) |
| At least one price in the visible text | 240 | 42.0% (38.0% to 46.1%) |
| An llms.txt file | 103 | 18.0% (15.1% to 21.4%) |
By kind of business
Kinds of business with at least 10 sites, largest first; the smaller ones are pooled in the last row. The last column adds the 95% interval.
| Kind of business | Sites | Cannot read | Call or email only | Labelled form | Standard system |
|---|---|---|---|---|---|
| Restaurants | 197 | 17.3% | 20.3% | 9.6% | 52.8% (45.8% to 59.6%) |
| Hotels | 127 | 22.0% | 10.2% | 29.1% | 38.6% (30.6% to 47.3%) |
| Gyms | 44 | 18.2% | 31.8% | 31.8% | 18.2% (9.5% to 32.0%) |
| Financial advisors | 31 | 19.4% | 54.8% | 25.8% | 0.0% (0.0% to 11.0%) |
| Dentists | 30 | 10.0% | 30.0% | 56.7% | 3.3% (0.6% to 16.7%) |
| Lawyers | 29 | 27.6% | 20.7% | 48.3% | 3.4% (0.6% to 17.2%) |
| Veterinarians | 27 | 18.5% | 40.7% | 22.2% | 18.5% (8.2% to 36.7%) |
| Car repair shops | 26 | 19.2% | 46.2% | 34.6% | 0.0% (0.0% to 12.9%) |
| Physiotherapists | 26 | 19.2% | 30.8% | 34.6% | 15.4% (6.2% to 33.5%) |
| Real estate agencies | 26 | 23.1% | 26.9% | 50.0% | 0.0% (0.0% to 12.9%) |
| Driving schools | 25 | 8.0% | 52.0% | 24.0% | 16.0% (6.4% to 34.7%) |
| Dermatologists | 20 | 10.0% | 50.0% | 35.0% | 5.0% (0.9% to 23.6%) |
| Architects | 18 | 5.6% | 88.9% | 5.6% | 0.0% (0.0% to 17.6%) |
| Hair salons | 17 | 5.9% | 35.3% | 11.8% | 47.1% (26.2% to 69.0%) |
| Electricians | 12 | 33.3% | 58.3% | 8.3% | 0.0% (0.0% to 24.2%) |
| Kinds of business with fewer than 10 sites | 39 | 12.8% | 43.6% | 35.9% | 7.7% (2.7% to 20.3%) |
By country
Largest samples first. Small samples have wide intervals, so read them as rough. The last column adds the 95% interval.
| Country | Sites | Cannot read | Call or email only | Labelled form | Standard system |
|---|---|---|---|---|---|
| Denmark | 67 | 16.4% | 31.3% | 11.9% | 40.3% (29.4% to 52.3%) |
| United States | 63 | 20.6% | 14.3% | 17.5% | 47.6% (35.8% to 59.7%) |
| Germany | 59 | 20.3% | 18.6% | 22.0% | 39.0% (27.6% to 51.7%) |
| Netherlands | 55 | 16.4% | 32.7% | 23.6% | 27.3% (17.3% to 40.2%) |
| Canada | 50 | 20.0% | 16.0% | 36.0% | 28.0% (17.5% to 41.7%) |
| France | 49 | 16.3% | 26.5% | 28.6% | 28.6% (17.8% to 42.4%) |
| Portugal | 43 | 11.6% | 30.2% | 25.6% | 32.6% (20.5% to 47.5%) |
| Belgium | 40 | 10.0% | 45.0% | 22.5% | 22.5% (12.3% to 37.5%) |
| Ireland | 38 | 13.2% | 23.7% | 52.6% | 10.5% (4.2% to 24.1%) |
| Australia | 37 | 27.0% | 21.6% | 8.1% | 43.2% (28.7% to 59.1%) |
| United Kingdom | 31 | 35.5% | 19.4% | 38.7% | 6.5% (1.8% to 20.7%) |
| Italy | 28 | 14.3% | 42.9% | 21.4% | 21.4% (10.2% to 39.5%) |
| Poland | 28 | 17.9% | 35.7% | 25.0% | 21.4% (10.2% to 39.5%) |
| Austria | 23 | 8.7% | 47.8% | 39.1% | 4.3% (0.8% to 21.0%) |
| Switzerland | 22 | 9.1% | 45.5% | 36.4% | 9.1% (2.5% to 27.8%) |
| Sweden | 18 | 22.2% | 44.4% | 33.3% | 0.0% (0.0% to 17.6%) |
| Spain | 17 | 17.6% | 52.9% | 23.5% | 5.9% (1.0% to 27.0%) |
| Norway | 14 | 7.1% | 57.1% | 21.4% | 14.3% (4.0% to 39.9%) |
| Mexico | 9 | 33.3% | 33.3% | 22.2% | 11.1% (2.0% to 43.5%) |
| Brazil | 3 | 33.3% | 33.3% | 0.0% | 33.3% (6.1% to 79.2%) |
How the study was made
- The sample
- Every own website that an earlier AIGROW Labs study linked to a business named by an AI assistant on an AIGROW Index page: 694 sites in 20 countries and 22 kinds of business, each counted once. Restaurants (197) and hotels (127) are a large part, so they are reported apart from the other trades (370). These are businesses the assistants already recommend, so the sample leans to well known places, and its shares describe them rather than all local businesses.
- The reader
- A small program that named itself openly as an AIGROW Labs research reader, sent one request per host at a time, waited at most 15 seconds for an answer and obeyed robots.txt under RFC 9309, treating a server error there as closed. On 5 November 2026 it read each home page and at most 3 linked pages about booking, appointments, a quote, a shop or contact. It ran no scripts, used no login, opened no checkout and submitted no form.
- The ladder
- Each site sits on the highest step reached on any page read. Cannot read: no usable page for an automated reader, from a dead name or a refusal to a page that is empty without scripts. Call or email only: no labelled form and no standard system. Labelled form: a request form whose fields all carry a name an agent can read; someone at the business still has to answer it. Standard system: one of 214 known booking, ordering and shop services or plugins, embedded or linked; shop software with a cart or product signal; or a schema.org action to reserve, order, buy or schedule that names a target.
- Statistics
- Every share comes with its count and a 95% Wilson score interval. Shares about robots.txt use the 668 sites whose robots.txt answered with a file or with a status meaning there is none. Shares about forms, links and markup use the 571 sites that could be read. 17 AI agent names were looked for in robots.txt.
- A lower bound
- The reader ran no scripts, and many booking widgets are drawn by a script after the page loads, so the top step is a lower bound. Counting every own booking form with date and guest fields, every cart or product signal and every empty booking frame raises it to 35.4% (32.0% to 39.1%). Sites that send an empty page and draw everything with scripts count as unreadable here, while an agent driving a full browser could read them.
- Few pages and a finite list
- Only the home page and up to 3 linked pages were read, so a booking page deeper in the site, or behind a menu built by a script, was missed. A service missing from the list is not counted, and a hotel group’s own booking engine counts at most as a labelled form.
- Found is not finished
- The ladder measures what an agent can find, not whether a booking succeeds. A standard system may still need an account, a payment card or a choice the agent cannot make, and a labelled form may get a reply days later.
- One reader, one moment
- Each site was read once, by one reader that named itself. Another user agent, network or hour would have met different refusals, as the comparison with the browser-like reading shows. Real AI agents differ in how they identify themselves and in whether they run a full browser.
- Small samples
- Several countries have fewer than 20 sites, and differences between countries partly reflect which trades the assistants named in each. Read the country and trade rows as rough.
- Competing interests
- The author founded AIGROW, which publishes the Index and sells services that aim to improve how businesses appear in AI answers. The study had no external funding.
Rankings from the AI Visibility Index
See which businesses and products ChatGPT, Claude, Gemini and Grok named for the categories and places below.
How to cite
Guglielmo Paesano, Can an AI agent book or buy on a local business website?, AIGROW Labs working paper, 5 November 2026, aigrow.gg/research/ai-agent-ready-sites-2026.
A working paper, not yet peer reviewed. The figures on this page are the paper’s.
The study read public pages only and names no business it studied. The per site file, which lists host names, stays with AIGROW Labs.
Can an assistant find you, and then book you?
The free scan asks ChatGPT, Claude, Gemini and Grok your own customers' questions, shows who they name instead of you, and checks whether their crawlers can read your site.
AIGROW Labs · 5 November 2026