Playbook 0: How we build websites in the AI era

We are a team of seven in Bruges, building websites for B2B service firms in Belgium and the Netherlands. Everything here is taken from our own project files in 2026.

Your website is the source AI reads about you

When a buyer asks ChatGPT, Gemini or Google's AI Mode which firm to call, the answer is assembled from web pages. An assistant can only repeat what your site states clearly, in text it can fetch. A firm whose services, process and proof sit in PDFs, sliders and one long homepage paragraph gets described vaguely, or gets left out in favour of a competitor whose pages answer the question directly.

Since we started measuring our own site in mid-August 2026, sessions from AI crawlers and agents have outnumbered sessions from people: 4,285 against 1,552 by 24 October. Two caveats pull in opposite directions. Nearly all of that machine traffic, 4,060 sessions, came from one crawler identifying itself as Meta's, which we could not verify. And our measurement only sees crawlers that run JavaScript, so most of the major AI crawlers are missing from the count. The firmer number is the human one: in the same period 24 people reached our site through a link in an answer from ChatGPT, Claude, Perplexity or Copilot, and eight of them landed on the same research paper. Assistants send buyers to specific pages, and they pick pages they could read.

The principles behind the order

1. Strategy is research. Every structural decision traces back to a search result, a client answer or a measured number, and the strategy report shows the evidence next to the decision. On one project the research reversed the language plan in our own proposal, because the buyers we were targeting searched in Dutch from inside the country.
2. One page, one job, one keyword. Each page gets a single job written as a sentence and at most one primary keyword. Two pages never compete for the same term.
3. Approve in order, then move on. The client approves the structure, then the homepage design, then the full design, then the built homepage, then the full site. Within a step we iterate until it is right, usually two or three rounds. Reopening an approved step goes through a change order, which is what lets us keep price and timeline fixed without capping revisions.
4. Build once, reuse without limit. Pages are built from templates and CMS collections, so a new service, case or article is a form to fill in. We price the number of unique designs a site needs, never the number of pages.
5. Measure before anything changes. We capture a baseline of the old site in week one, because once redirects go live the before picture cannot be reconstructed. After launch, every piece of work is measured against the number it was meant to move.
6. Commit only to what the results page supports. We split target keywords into a group we commit to and a group we work toward, after opening the actual Google results for each one. Playbook 2 shows how.

What we choose to leave out

We present one design direction, applied fully to the homepage, because the strategy phase has already settled positioning, audience and structure. We do not build webshops: optimising a store is a different discipline, and we refer that work to partners. A second language that carries its own keywords is written from its own research; a language that only needs to exist is translated with AI and reviewed by a person on the pages that convert.

The four phases, the approval that closes each one and the usual duration are drawn below. A typical project runs six to ten weeks. Late input moves the launch date day for day, and we say so in the proposal.

Within a phase we iterate until the client is satisfied; reopening a phase that was already approved goes through a change order.