Fifty Amsterdam websites, read the way a machine reads them

This is a joint piece of fieldwork by ebya.nl and Landingpeak. Neither of us was hired to write it and nobody in the table below asked to be measured. It exists because the two of us kept having the same argument with business owners in Amsterdam — that the website is “fine”, that it “looks good”, that the phone rings anyway — and arguments are cheap next to a number you can check yourself.

Here is the thing that changed and that almost nobody has priced in yet. A customer with a burst pipe at eleven at night no longer opens ten blue links. They ask an assistant — ChatGPT, Gemini, the AI block above Google’s results, Copilot inside Windows — “spoed loodgieter Amsterdam, wie kan nu komen en wat kost het?” and they get back two or three names and a rough figure. If your business is not one of the names, you were not outbid. You were not in the room.

Being in the room is a different job from ranking. Ranking asks whether Google can find you. Our proxy asks whether an AI system can read your page, reconcile its identity, find something specific enough to quote, and crawl the page at all. Those four questions are the four groups every site below is scored on. This is a diagnostic model, not a measurement of any AI company’s private ranking system.

What the fifty look like

The one finding that matters

Every study like this hopes for a villain, and usually the villain is technical: broken markup, blocked crawlers, JavaScript that hides the content. That is not what we found. Technically these fifty are mostly fine. The access subscore averages 82.6/100, the pages weigh a median of 138 KB, and the median server answers in 271 milliseconds.

All 49 readable sites — every single one — fail the check for a quotable number. Amsterdam’s trades are legible and unquotable at the same time.

Read that again in the shape of the customer’s question. Somebody types “wat kost een tandarts controle in Amsterdam?” or “prijs elektricien Amsterdam per uur”. A useful answer needs a figure and a source. Across these 49 readable pages, we found no published rate that could be attributed to the business. That leaves directories, price-comparison sites and national chains with published rates as easier sources. The local business may have earned the customer’s attention through reviews, then surrendered the most quotable part of the answer.

The same pattern repeats one level down. 65% state no response time — no “binnen 24 uur”, no “same day”, no opening hours in text a machine can lift. 49% answer no questions in a clear question-and-answer format. 55% carry no review signal in the page text at all, despite most of them having real Google reviews sitting one click away. The reputation exists. It just is not written on the page we measured.

Four questions, four scores

Each site gets a score in four groups. Splitting them apart is what makes the diagnosis useful, because the fifty are not uniformly bad — they are excellent at two of these and dreadful at the other two, and that is a much cheaper problem to fix than it sounds.

The table

Sorted best to worst. “Faults” lists the checks each site fails outright, most costly first; a warm-coloured chip is a warning rather than a failure. Server time is measured machine-to-machine and is not the same as what a phone on 4G experiences — treat it as the floor, not the ceiling. Words is the visible body copy on the homepage.

Every row was fetched live on 2026-09-04. Scores move when sites change; the dataset behind this table is stored alongside the page so the figures can be re-checked rather than taken on trust.

By trade

Splitting the sample by niche is where it stops being a curiosity and starts being useful, because the gap between trades is bigger than the gap between the best and worst site inside most of them.

Plumbing leads for an unglamorous reason: emergency work forces a business to say when it can come, and saying when you can come is one of the checks. Physiotherapy trails for the mirror-image reason — a physio site is usually built to look calm and reassuring, which in practice means beautiful photographs, very few nouns, and nothing a model can lift. A clinic with 284 reviews and a page that says almost nothing is a clinic that will not be recommended by a machine, no matter how good the treatment is.

What we would fix first, in order

None of the top five costs money to say. They cost a decision about what to publish, which is why almost nobody does them and why doing them moves a score more than any technical work on the list.

1. Put one real number on the homepage. A call-out fee, an hourly rate, a from-price, a consultation price. “Vanaf €79 voorrijkosten” is a sentence a model can quote; “neem contact op voor een offerte” is a sentence nobody can quote, including a human.
2. Say when you answer. “Binnen 2 uur bij spoed, 24/7 bereikbaar” answers the second half of every emergency question and 32 of the 49 sites here answer it nowhere.
3. Name the ground you cover. Not “Amsterdam en omstreken” — the districts and towns, written out: Centrum, Noord, Zuidoost, Nieuw-West, Amstelveen, Diemen. 36 sites never do this.
4. Answer ten questions in question-and-answer form. The literal questions customers ask on the phone, with the literal answers. This is the format assistants quote most reliably, and it is also the cheapest page anyone will ever write.
5. Mark the business up as a LocalBusiness. Name, address, phone, hours, area, price range in structured data, so the machine does not have to guess what kind of thing it is reading. 13 sites here carry no structured data at all.

Straight answers