An AI agent can book or buy on your website when it finds a standard system: a known booking or ordering service linked from your pages, shop software with a working cart, or markup that declares a booking action with a target. AIGROW Labs read 694 local business sites in 20 countries, and 27.1% offered one.
Run a local business? AI Local Fix sets up your listings, Google profile and markup, then counts the answers that name you.
- An AI agent finishes a booking or an order when your site offers a standard system it recognises: a known booking or ordering service, a shop with a working cart, or markup that declares a booking action with a target.
- In a study of 694 local business sites in 20 countries, 27.1% offered one: 52.8% of restaurants, 38.6% of hotels and 9.5% of the other trades.
- Nearly half of the sites (47.4%) gave an agent nothing to act on: it could not read them, or it found at best a phone number or an email address.
- Bot defences shut out more agents than robots.txt did: 8.9% of all sites refused the reader, while 3.7% of those whose robots.txt answered closed the home page to an AI agent by name.
- The free AI agent readiness checker reads your site the same way and gives four plain answers, with no account.
What an AI agent does when someone asks it to book
Ask an AI assistant with agent features to book a table for four on Friday, and it does what a person in a hurry would do. It picks a business, opens its website, reads the page and looks for a way to act. If it finds one it can use, it books.
If it does not, the most it can do is hand back a phone number.
The agent does not see your site the way a customer does. It may not run every script, it cannot guess what an unnamed box on a form is for, and the security service in front of your site may turn it away before it reads a word. What it can do with your site falls on one of four steps.
| Step | What the agent finds | What your customer gets |
|---|---|---|
| 0. Cannot read the site | A refusal, a challenge page or a page with no text | Nothing from your site |
| 1. Read only | Your text, and at best a phone number or an email address | A number to call or an address to write to |
| 2. Labelled request form | A contact, quote or booking request form whose fields carry readable names | A request, which someone at your business still has to answer |
| 3. Standard system | A known booking or ordering service, a shop with a working cart, or a declared booking action | A booking or an order, if the system lets the agent finish |
Only the top step lets the agent complete the job by itself. The second step gets a request to you. Steps 0 and 1 leave the work with your customer, who may simply ask the assistant for another business.
What a study of 694 local business sites found
On 5 November 2026, AIGROW Labs read the own websites of 694 local businesses in 20 countries and 22 trades. They are the businesses AI assistants named in an earlier AIGROW Labs study, so the sample leans towards well known places that an assistant is likely to recommend.
A small program read each site the way a simple agent would. It named itself openly and obeyed robots.txt. It read the home page and up to 3 linked pages about booking, a shop or contact. It ran no scripts, logged into nothing and submitted nothing. Then it placed each site on the four steps.
| Step | Sites | Share of 694 |
|---|---|---|
| 0. Cannot read the site | 123 | 17.7% |
| 1. Read only | 206 | 29.7% |
| 2. Labelled request form | 177 | 25.5% |
| 3. Standard system | 188 | 27.1% |
Most of the top step came from a known booking or ordering service: 168 of the 188 sites (89.4%) used one, 115 of them inside a page and 53 through a link only. Shop software with a working cart carried 36 of the 188 (19.1%), and 13 sites declared a booking or ordering action in their markup.
The other side is bigger. 329 sites (47.4%) gave an agent nothing to act on: 123 could not be read at all, and 206 could be read but held no usable form and no system. On 156 of those 206 (75.7%), the agent could at least find a phone number, an email address or a WhatsApp link to pass back.
Since the reader ran no scripts, the 27.1% is a floor. If you also count each site's own booking form with date and guest fields, each cart or product signal and each booking frame left empty for a script to fill, the share that offers some way to book or buy rises to 35.4%. Even then, most sites expect a person to call, write or send a request.
What it found for your kind of business
The trades are far apart. Restaurants are the most ready, hotels come next, and for most other local businesses the top step is rare.
| Sector | Sites | Cannot read | Read only | Labelled form | Standard system |
|---|---|---|---|---|---|
| Restaurants | 197 | 17.3% | 20.3% | 9.6% | 52.8% |
| Hotels | 127 | 22.0% | 10.2% | 29.1% | 38.6% |
| Other trades | 370 | 16.5% | 41.4% | 32.7% | 9.5% |
Restaurants
Just over half (104 of 197) already use a booking or ordering service an agent can recognise. Still, 37.6% give an agent nothing to act on, and 17.3% could not be read at all. If your restaurant takes bookings by phone only, an agent asked for a table can do no more than pass your number on.
Hotels
49 of 127 hotels (38.6%) reached the top step. Some large hotel groups run their own booking engine, which the study's list of services did not include, so they count as a labelled form or lower. The same goes for any booking page that only a script reveals, or that your home page does not link to.
Clinics and practices
The top step is rare here: 1 of 30 dentists, 1 of 20 dermatologists, 4 of 26 physiotherapists and 5 of 27 veterinarians. What these sites offer most is a request form. 17 of 30 dentists (56.7%) had a labelled one, so an agent can ask for an appointment, but someone at the practice still has to reply.
Half of the dermatologists (10 of 20) could be read but offered no form and no booking system.
Salons, gyms and other trades
Hair salons did best outside food and travel, with 8 of 17 (47.1%) on the top step, though 17 sites is a small sample. Gyms reached 18.2% (8 of 44), driving schools 16.0% (4 of 25) and lawyers 1 of 29. No site on the top step turned up among financial advisors (31 sites), car repair shops (26), real estate agencies (26), architects (18) or electricians (12).
Shops
The study kept no separate row for shops, but it counted shop software with a working cart as a standard system: a link to a cart or a product page, or an add to cart button, showing the shop is in use. That is how 36 of the 188 top step sites got there. A shop platform with a visible cart is something an agent can recognise.
A catalogue that asks people to call to order is not.
By country
| Country | Sites | Standard system |
|---|---|---|
| United States | 63 | 47.6% |
| Australia | 37 | 43.2% |
| Denmark | 67 | 40.3% |
| Germany | 59 | 39.0% |
| Portugal | 43 | 32.6% |
| France | 49 | 28.6% |
| Canada | 50 | 28.0% |
| Netherlands | 55 | 27.3% |
| Belgium | 40 | 22.5% |
| Italy | 28 | 21.4% |
| Poland | 28 | 21.4% |
| Norway | 14 | 14.3% |
| Mexico | 9 | 11.1% |
| Ireland | 38 | 10.5% |
| Switzerland | 22 | 9.1% |
| United Kingdom | 31 | 6.5% |
| Spain | 17 | 5.9% |
| Austria | 23 | 4.3% |
| Sweden | 18 | 0.0% |
Brazil had only 3 sites and is left out of this table. Country shares partly reflect which trades the assistants named in each country, and a country or trade with 20 or 30 sites gives a rough reading, not a precise rate. The full tables are on the study page.
What lets an agent finish the job
Each point below comes from what the study recorded on the sites that reached the top step, or from what stopped the others.
Let the agent in
Few owners block AI agents on purpose. Of the 668 sites whose robots.txt answered, 25 (3.7%) closed the home page to an AI agent by name, and every one of them closed it to a crawler that collects training data. Only 4 (0.6%) closed it to the agents that open a page because a person asked, such as ChatGPT-User, Claude-User and Perplexity-User.
The door that does close is the firewall. 62 sites (8.9% of all) answered the reader with an error such as 403 or 429, or with a bot challenge page. Those refusals were not stable.
An earlier AIGROW Labs study read the same home pages on the same day with a normal browser reader: of the 556 sites that served it, 27 (4.9%) turned the study's reader away, and of the 60 that refused it, 21 (35.0%) let the study's reader in.
If your host or a security service screens visitors, ask it which automated visitors it challenges, and let in the agents you want to reach you. In robots.txt, keep the three kinds of AI robot apart: crawlers that collect training data, crawlers that build an AI search index, and agents that fetch a page for a person. You can close the first kind and still welcome the third.
Put your text in the page
20 sites (2.9%) sent an empty page and drew everything with scripts, so a reader that runs none saw no text at all. Ask your developer whether your opening hours, services and booking link are in the page before any script runs.
Link your booking or shop system plainly
Put a plain link to your booking service or shop on the home page, or one click from it, with words such as Book, Reserve or Order. The study read the home page and up to 3 linked pages, so a booking page deeper in the site, or behind a menu built by a script, was missed. 28 readable sites (4.9%) had a booking frame left empty for a script to fill, which a simple agent cannot use.
Give every form field a name
If a request form is your best offer, give every field a readable name: a label tied to the field, an accessible name, a title or a placeholder. 58 of the 571 readable sites (10.2%) had only forms whose fields carried no readable name, so an agent would have to guess which box takes the name, the date or the message. A form that asks for a date and a number of guests counts as your own booking form; 57 sites (10.0%) had one.
Write the facts down in markup
Structured data tells an agent what you are and what it can do on your site. 78.8% of readable sites carried some schema.org markup, but only 25.7% described the business as a local business type, 13.0% gave opening hours in markup and 2.3% (13 sites) said whether they take reservations. Just 13 sites (2.3%) declared an action to reserve, order, buy or schedule with a target an agent can follow.
A short example for a restaurant:
{
"@context": "https://schema.org",
"@type": "Restaurant",
"name": "Example Bistro",
"url": "https://example.com",
"telephone": "+1 555 0100",
"openingHours": "Tu-Su 12:00-22:00",
"acceptsReservations": true,
"potentialAction": {
"@type": "ReserveAction",
"target": "https://example.com/book"
}
}A clinic would use its own type, such as Dentist, and a schedule action where it takes appointments. The target must be the real address of your booking page.
Keep the phone and email links
Across the 571 readable sites, 62.5% had a tappable phone link and 57.4% an email link. They do not let an agent finish the job, but they give your customer a way through when nothing else works.
How to check your own site
The free AI agent readiness checker reads your site the way the study did: robots.txt, your home page and up to 3 pages it links to, with no scripts, no login and nothing submitted. It needs no account, and it answers four questions:
- Can an agent book or buy here through a standard system?
- Does robots.txt close your home page to an AI agent by name?
- Did a bot defence refuse the reader?
- Is there structured data an agent can act on?
Each answer shows what was found and on which page, next to the same figure from the study. Start with the first answer that comes back red. If the checker cannot read your site at all, begin with the firewall, because nothing else counts until an agent gets in.
The checker looks at one moment and one reader. If it shows a refusal, run it again later before you change anything, since the study found that refusals shift from one reader and one hour to the next.
Where AI Local Fix fits
An agent books the business the assistant picked. Before any of the above pays off, ChatGPT, Claude, Gemini or Grok has to name you when someone asks for a dentist or a table nearby. That part is the job of AI Local Fix.
It claims, corrects or creates up to 15 listings on the directories assistants cite for your trade and country, completes your Google Business Profile and puts the same facts on Bing Places and Apple Business Connect. It adds LocalBusiness markup to up to 10 of your pages, writes 3 answer pages from the questions your customers ask, and prepares a reviews kit.
It also measures. Before the work starts, at day 30 and at day 90, it asks ChatGPT, Claude, Gemini and Grok 10 questions for your city and counts how often they name you next to your 3 closest rivals. If the day 90 count names you no more often than the first, you get a full refund.
Nobody can promise a ranking in AI answers, and AI Local Fix does not.
It does not install a booking system or a shop. If the checker shows you have none, pick one that suits your trade and link it from your home page; AI Local Fix then works on getting your business named, and its markup tells an agent what you are and when you are open.
What the study cannot tell you
- The reader ran no scripts, so a booking tool that loads by script was missed. The 27.1% is a floor, not an exact count.
- It read the home page and up to 3 linked pages, and it counted only services on a written list of 214. A hotel group's own booking engine counts at most as a labelled form.
- Finding a system is not the same as finishing a booking. A system may still need an account, a payment card or a choice the agent cannot make, and a request form may get a reply days later.
- Each site was read once, by one reader, on one day. Another reader or another hour would have met different refusals.
- The sample leans towards businesses that assistants already recommend, and several trades and countries have few sites.
- AIGROW sells services that touch these findings, AI Local Fix among them. The study is a working paper, not peer reviewed, and its method and limits are on the study page.
Rather have this done for you?
AI Local Fix does the groundwork once: your Google Business Profile, up to 15 listings, the markup on your pages and a reviews kit. It counts how often four assistants name you on 10 questions for your city before the work, at day 30 and at day 90.
If the day 90 count names you no more often than the first, you get the full price back. Ask within 14 days of the day 90 report. No ranking and no specific AI answer is promised.
Frequently Asked Questions
