A customer with a 2018 Mercedes wants to know whether wireless CarPlay can be retrofitted to it, and what that costs. They do not browse. They type the question, more or less as a sentence, and they buy from whoever answers it.
Until fairly recently, answering it meant ranking for it. Now a growing share of those questions never reach a results page at all. Someone asks an assistant, it reads a handful of sources, and it produces an answer with two or three citations. If you are not one of them, you were not in the room.
We built os-cartech.de for a retrofit workshop in Viersen — ambient lighting, CarPlay, reversing cameras, the kind of work that is meant to look like it came from the factory. This is what we changed to make the site legible to the systems now doing the answering, and which of those changes were worth the effort.
The three things that actually mattered
Most of what gets written about this is speculative. These three are not — they are mechanical properties of how the crawlers work, and you can verify each one yourself in about ten minutes.
AI crawlers do not run JavaScript
This is the big one, and it invalidates a lot of otherwise-reasonable architecture.
A search engine crawler will, eventually, render your page. The crawlers feeding answer engines mostly do not. They fetch the HTML and read what is in it. If your content arrives after a script runs — a client-side fetch, a framework that hydrates an empty shell, a tabbed interface that only populates on click — then as far as those systems are concerned, the page is blank.
We built every page on the site to be complete in its initial HTML. Not "server-rendered with client hydration for interactivity" as a nice-to-have, but a hard rule: if you disable JavaScript entirely, every word on the site is still there.
You can check your own site in one command. Fetch the page without executing anything and look at what comes back. If the interesting part is missing, nothing else in this post will help you.
They quote passages, not pages
An answer engine is not deciding which page is best. It is looking for a chunk of text that answers the question, which it can lift more or less intact.
That changes how you write. The journalistic instinct — set up the context, build the argument, arrive at the conclusion — produces text where the answer is in the last paragraph and is meaningless without the first four. That is unquotable. A passage that begins "It depends on several factors" gives the machine nothing to take.
So every page on the site is organised as questions, with the answer in the first sentence underneath, and the qualification after that. Not because it games anything, but because it turns out to be how people want to read this material anyway. Someone who wants to know whether their car supports wireless CarPlay does not want three paragraphs of preamble either.
The practical shape: a real heading that is the actual question, a direct answer in one sentence, then 100 to 150 words of the detail that makes the answer trustworthy. Long enough to be worth quoting, short enough to be quoted whole.
They can only read you if you let them
This is the one almost nobody does deliberately.
Your robots file either permits the AI crawlers or it does not, and most sites have never thought about it. Some block them through an inherited rule nobody revisited. Some permit them by default and have no idea. Either way it is an accident.
We named them explicitly — the crawlers behind the major assistants and answer engines, each given permission in its own line. That is not a technical trick, it is a decision, and it should be made on purpose rather than inherited from a template.
Worth being clear about the trade-off, because there is one: allowing these crawlers means your content can be summarised without anyone visiting your site. For a business selling a considered local service, that is fine — the summary carries the phone number and the town. For a business whose product is the content, it is a genuinely harder call, and anyone who tells you otherwise is selling something.
What we built, structurally
The site is twenty-three pages. That number is not arbitrary and it is not padding.
There are six services, four car brands, and six guides answering questions people actually ask — whether a dashcam is legal here, whether a retrofit voids the warranty, what ambient lighting costs, whether wireless CarPlay is worth it over wired. Plus the ordinary pages a business needs.
Each of those pages answers one question properly. The alternative — one long services page covering everything — is a page that answers every question badly and can be quoted for none of them.
The guides are the part that does the work. They are commercial-intent questions with honest answers, including the answers that lose a sale. The warranty guide says plainly when a retrofit would affect a warranty. The dashcam guide explains the legal position without pretending it is simpler than it is. A workshop that tells you when not to buy is more credible than one that does not, and — usefully — that is also the kind of passage a machine is happy to quote, because it reads as information rather than as marketing.
The structured data question
We added structured data describing the business: what it is, where it is, when it opens, which towns it serves, what it does. Not decoration on the page, but statements of fact in a format designed to be read literally.
This is worth doing, and it is worth being honest about why. The direct search-result benefits have narrowed considerably — Google retired FAQ rich results for essentially everyone in May, and several other formats have gone the same way over the past few years. If you are adding structured data because you saw a screenshot of a fancy search result in 2021, you may be disappointed.
The reason to do it now is different: it lets a machine state a fact about the business rather than infer one. "Open until 18:00" is better than a guess drawn from prose. That is a smaller, duller benefit than the one people were sold, and it is still worth the afternoon.
What we did not do
We did not write an llms.txt file and expect anything from it. We ship one — it costs nothing and it makes the site legible to anything that reads it — but there is no evidence it affects citation today, and anyone presenting it as a ranking factor is guessing.
We did not add a blog on a content calendar. A workshop does not need to publish weekly. It needs six pages that answer the six questions its customers actually ask, written properly once.
We did not chase keyword volume. The questions this business gets asked are low-volume and high-intent. Ten people asking whether their specific car supports a specific retrofit are worth more than a thousand people reading about car interiors in general, and the ten are considerably easier to reach.
What we would do differently
The brand pages and the guide pages overlap more than they should. If we started again we would write the guides first and let the brand pages point at them, rather than building both in parallel and then reconciling. That is a day of work we did not need to spend.
The uncomfortable part
None of this is measurable in the way SEO was measurable.
You can watch your rankings. You cannot watch your citations — there is no console that tells you an assistant recommended you to someone last Tuesday. The tools that claim to are sampling a handful of prompts and extrapolating hard.
So the honest position is this: the changes above are cheap, they are all things that make the site better for humans regardless, and they are mechanically necessary if you want to be readable by systems that increasingly sit between a business and its customers. That is a good enough reason to do them. "We will show you a dashboard proving it worked" is not something we can offer yet, and we would be careful with anyone who says they can.
The workshop's site is faster, clearer, and answers more questions than it did. That part we can show you.