G-Lab field study · Amsterdam · July 2026
Restaurants are visible to people. Their booking systems are not ready for agents.
We took 200 Amsterdam restaurants and walked the path a guest walks: find the place, understand it, check whether it fits, find a free table, book it, get a confirmation. Then we walked the same path as an agent.
Being online is no longer the same as being bookable.
A restaurant can rank well, hold a thousand reviews, run a beautiful site and show a large Reserve button, and still be impossible for an assistant to book. That gap is what this report is about.
The booking journey got shorter
Discovery used to start with a search result and end with a person clicking Reserve. Six steps, all performed by the same pair of hands.
- Find the restaurant.
- Understand what it offers.
- Check whether it fits the request.
- Find availability.
- Book a table.
- Receive confirmation.
A guest can now say this instead:
Find a quiet seafood place near the canals, outside seating, table for two, Friday at seven.
Something has to interpret that, compare restaurants, check real availability and finish the reservation. The restaurant is no longer judged only by a person. It is also judged by whatever is acting on that person’s behalf.
What we tested
200 restaurants, four layers, one identical pass
The sample was drawn at random from 1,631 operational Amsterdam restaurants with a fixed seed, so nothing was hand-picked. Every venue got the same checks, in the same conditions, on the same day. We tested only the public path a customer would use. We never opened an admin system, and no reservation was ever submitted.
Online presence
Could we reliably identify the restaurant, its official website, location, hours and reservation path?
Customer actions
Could a person find the menu, call, get directions or start a reservation without hunting for it?
AI readiness
Were the essential facts stated in a form a machine can read, or scattered across Instagram, a PDF and three pages?
Agent-ready booking
Could an agent locate the reservation system, check a specific date and party size, choose a time and reach the point where only the guest's own details remained?
Measured, Amsterdam 2026
Every step loses half the restaurants
We ran 200 Amsterdam restaurants through the same checks, in the same conditions, on the same day. This is how far each one gets before an AI stops being able to do anything.
↓ 77 drop out here
↓ 50 drop out here
↓ 16 drop out here
↓ 16 drop out here
G-Guest sits on this last rung. Not one restaurant in the sample publishes a booking interface an assistant can call. Ours does, and independent agents have used it to create confirmed reservations.
n=200 · random sample, fixed seed · Amsterdam municipality · July 2026 · no reservation was ever submitted
The results
41% had a booking path a machine could see
The rest ran on contact forms, phone numbers, Instagram messages, third-party profiles or booking links buried where nothing could find them. 37.5% offered no route at all that we could detect. Another 5.5% asked people to call or message.
For a person this is friction. They squint, scroll, try the other menu item, find it. For an agent it simply ends the task.
The information usually existed. The structure did not.
The median score across the sample was 52 out of 100. Most restaurants published their facts, then scattered them across Google, Instagram, a PDF menu and three separate pages. Hours in one place, a different set in another. A Menu link pointing at last spring.
An assistant should not have to guess whether a place is open, whether it takes reservations, or which menu is the real one. When the facts disagree, recommending a different restaurant is the safer answer. That is the quiet way a venue loses.
33% sat behind a third-party widget, and it did not save them
Booking platforms solve a real operational problem. They also introduce a layer between the restaurant and whoever is trying to reach it. The button is visible to a person while the availability, the fields and the confirmation live inside an iframe or a script.
Of the 35 venues running a paid booking product, 29 could still not be completed. That is 83%. The usual cause was a widget that renders beautifully and exposes no readable fields, followed by CAPTCHA walls, cookie walls and login walls. Paying for booking software does not make a restaurant bookable by software.
8% could be taken to the final step. None could be booked without a browser.
This is the number that matters. An agent driving a real browser reached the point where only the guest’s own name and phone remained at 16 restaurants out of 200. We stopped there every time, on purpose.
Then we checked something else: whether any restaurant publishes a booking interface an assistant could call directly, without pretending to be a person with a mouse. Across 163 working domains, not one did. Nine served a machine-readable file, and all nine were generated automatically by an SEO plugin. They described the restaurant. None of them contained a way to book it.
The failure almost never happened at discovery. It happened at action.
A booking button is not a booking interface
Most restaurant sites were built on a visual assumption. A person sees a button, recognises a calendar, understands the widget, knows what to do next. All of that understanding happens in the person, not on the page.
Software needs the steps to exist as actual steps: check availability, choose a slot, give the guest details, create the booking, return a reference, confirm it plainly.
This does not mean removing the visual experience. It means both audiences reaching the same underlying action. One restaurant. One availability. Two ways in.
You will never get a report about the booking you lost
Nobody calls to say their assistant could not use your reservation widget. There is no failed payment, no abandoned checkout, no lead sitting in a CRM waiting to be chased.The assistant offers a different restaurant and the conversation moves on.
This is why the problem is almost impossible to notice from inside the business. Everything looks fine. People are still booking, because people can still click. A second channel is quietly being tested against you, and it does not file complaints.
Amsterdam is the warning, not the exception
This is one of Europe’s most digitally mature restaurant markets. Strong concepts, polished branding, deep review profiles, established booking platforms. If the gap is here, it is not a problem of weak websites in slow towns. It is structural.
The wider numbers point the same way. 45% of consumers now use AI to find a local business, up from 6% a year earlier (BrightLocal, 2026). In Uberall’s 2026 benchmark of quick-service restaurant locations, 83% did not appear in AI-generated recommendations at all. That figure is not ours and not Amsterdam-specific, but it describes the same shift from a different angle.
Discovery moved first. Action follows discovery. It always has.
What ready looks like
Four layers, and none of them is a redesign
A clear public identity
One consistent name, address, hours, cuisine, menu and policy. Not four versions across four places.
A website a machine can read
Business facts stated plainly and structurally, not implied by a photograph of a chalkboard.
A direct reservation path
Availability and booking that do not live entirely inside a closed visual widget.
A confirmation that says so
A booking returns an explicit status and a reference the guest and the assistant can both read.
Together these turn a website from a brochure into something that can be operated.
For ten years the goal was to be found
Rankings, Maps, reviews, social, a site worth looking at. All of it still counts. But being found stopped being the finish line, and the new question is short:
Can whatever found the restaurant also finish the booking?
A restaurant that can be understood but not booked leaves the journey unfinished. One that can be discovered, judged and booked stays usable as the way people search keeps changing.
What we built after measuring this
One reservation system with two ways in
We did not set out to build booking software. We set out to measure a gap, found that nothing in the sample cleared the last step, and then had to answer the obvious question: what would clearing it actually look like?
G-Guest gives a restaurant the familiar booking experience for guests, and a structured path an assistant can discover and use. Same availability, same inbox, same guest list. The restaurant keeps the relationship, and stops being invisible to whatever is trying to book it.
Independent agents have used it: OpenAI’s Codex, Claude, and an agent built on the open-source Hermes have all created confirmed reservations through the public interface, with no account and no key.
Where does your restaurant stop?
The diagnostic walks the same path this study did, on your own site, and shows you the rung you fall off. Four scores: online presence, customer actions, AI readiness and agent-ready booking.
Free, about a minute. No credit card.
Research notes
- Sample: 200 venues drawn at random with a fixed seed from 1,631 operational Amsterdam restaurants in Google Places. Collected on one day in July 2026. Availability, sites and booking providers change; this is a snapshot.
- Two independent engines ran all 200. They agreed on 180. Every disagreement was reviewed by hand, including four venues that reached a form which turned out not to be a table booking at all: a food-ordering checkout, a private-events enquiry, a campsite stay.
- A failed agent test does not mean customers cannot book. It means the public booking path could not be completed reliably without a person filling the gaps.
- We measured whether the path is traversable, which is an upper bound on what a real assistant achieves, never a floor. Method and version: Agent-Ready Booking Score v1.2.
- This study looks at digital accessibility and booking infrastructure. It says nothing about the food, the service or the hospitality, and most of these restaurants are very good at all three.
- No venue is named as a low scorer. Any business can request its own full result privately, or ask to be excluded, at hello@g-lab.studio.