Research · 6 September 2026

Being found isn't being booked

Thirty-four Amsterdam restaurants started writing for AI this summer. Not one made itself bookable by it.

Since April, if you ask Google to find you a table for Friday, something happens that did not happen a year ago. AI Mode does not hand you ten blue links. It reaches into booking platforms, pulls live availability, and shows you actual open times inside the answer. Worldwide, no subscription. You still tap to confirm, so nothing here is fully automatic.

But look at what already happened before that tap.

The shortlist was built. The times were checked. The choice was all but made. And nobody visited a website to do any of it.

Which leaves exactly one question for a restaurant, a salon or a clinic: was your availability readable enough to be in that list?

I got this wrong myself, three days before writing this. I had read Google’s announcement about food ordering in Ask Maps, seen that tables were not in it, and concluded that agentic table booking did not exist yet. It does, and it has since April, in a different product with a different announcement. Two days later I had to go back and correct four emails that had already gone out to editors. So if the shift feels hard to keep up with, that is not you being slow.

We measured it, and then measured it again

In July we took 200 Amsterdam restaurants, drawn at random from 1,631 in the city, and went through each one the way a guest would. First as a person. Then as an agent driving a real browser: find the booking path, open it, read the fields, fill date and time and party size, and stop at the moment the flow asks for the guest’s own details. We never submitted a reservation. Stopping there is the measurement: “reached the point where only the guest’s name and phone remained” is exactly what “an assistant could have finished this” means, without leaving a phantom table in someone’s book.

On 79 an agent could find a booking path at all. On 32 it could read the actual date and time fields. And across all 163 working sites, not one published a booking interface an assistant could call directly, without pretending to be a person with a mouse.

July was two months ago, so at the start of September we ran the whole sample again.

Nothing improved. Twelve reached the final step instead of sixteen. Seven sites no longer load at all. The number of venues exposing a callable booking interface was still zero.

The one thing that did change

On 6 September we checked every one of those 163 domains for the files that make a business legible to software: an MCP manifest, an OpenAPI description, an AI plugin descriptor, an llms.txt.

Thirty-four of them now publish llms.txt.

That file did not exist on those sites in July. It is a plain text document that explains a business to a language model: who we are, what we serve, when we are open. One in five restaurants in the sample noticed that a new kind of reader had arrived, and wrote something for it.

Not one of them published a way for that reader to book a table.

That gap is the whole story. The market heard “AI is coming” and answered with prose. It is the right instinct and the wrong layer. An assistant that can read your llms.txt still cannot reserve anything; it can only describe you accurately while sending the guest somewhere the job can actually be finished.

Nothing is broken, which is why nobody fixes it

None of those 163 websites is bad. They load fast, they look good, they turn visitors into customers. They were built for a web where being found was the hard part, and they are very good at being found.

A restaurant can rank first, have a thousand reviews, and still be absent at the exact moment a guest is ready to book, because its availability exists only as pixels inside a widget that renders beautifully and exposes nothing.

Paying for booking software does not solve it either. Of the 43 venues where we could identify the booking product behind the button, at 31 an agent still could not get a booking finished. That is 72 percent. And it is not any one vendor’s fault: we tested two venues on the same well-known system on the same day, and one blocked the agent at a cookie banner while the other let it reach confirmation. Same software, different configuration, different outcome. Which means it is fixable. Somebody just has to measure it.

There is a subtler version of the same problem. At 26 of the 200, the agent ran into the venue’s own protection: an anti-bot check, a captcha, a device check. Those tools are there for real reasons, and they were installed on the advice of the people who build restaurant websites. They also do not distinguish a scraper from a guest’s assistant. This is not an argument for removing protection. A consent banner is not security at all, since a scraper never renders it. A captcha is real security, but it is usually applied across the whole site, including the part that only reads opening hours and free tables. Those are public facts, given out over the phone all day. Protect what can be damaged, not what can be asked.

Why the owner never finds out

The cruel part is that none of this generates a signal.

There is no error message when an assistant gives up. No missed call. No abandoned cart, no lost lead sitting in a CRM waiting to be followed up. The assistant simply assembles its list out of what it can read, and moves on to a venue where it can finish.

Twenty years of local SEO trained an entire industry to chase being found, and built good instincts for measuring it: rankings, impressions, clicks. Not one of those measures picks this up. You can be winning every one of them and still be invisible in the half-second when the decision is actually made.

Being found isn’t being booked

Being in the answer gets you considered. Being usable gets you the table.

Discovery is close to solved: assistants are good at finding a suitable restaurant and explaining why it fits. It is the moment after the recommendation that breaks, and it breaks quietly, and it breaks for businesses that have done nothing wrong.

If you run a place that takes bookings, the question worth asking this month has changed. Not how you rank. Whether the thing that found you can also finish the job.

You can check that on your own site in about a minute, and we would rather you tested it yourself than took our word for it.