EN
On 17 July we discovered that Google didn't know tech.admix.nl. Not ‘hard to find’ — absent. By now 12 of the 13 pages are in the index, the first target terms have appeared in the search results and the first AI assistants have come by. This is exactly what we do for clients, so we're showing the numbers — including the conclusions we had to retract along the way. And that one page Google still doesn't know? That's the article you're reading right now. More on that below.
tech.admix.nl went live on 9 July. A week later we did what we do first for every client: ask the Search Console API what Google thinks of the site.
The answer was the same for every single URL: URL is unknown to Google. Impressions over 90 days: 0.
That is a different problem from a poor position. With a poor position you can optimise. With ‘unknown’ the page doesn't exist — you can rewrite the copy until you're blue in the face, nothing changes, because nobody is reading it.
Three causes, all three boring:
Submit the sitemap. Via the API, on the URL that itself returns a 200. That last part is not a detail: on our main site, the sitemap was registered on the variant without www, which redirects to the variant with it. Google doesn't follow that redirect and reported 21 errors — since 29 January. Six months of a dead signal, while the sitemap itself was fine the whole time. Resubmitted on the correct URL: fetched within a minute, 0 errors.
On tech the effect was immediately measurable. Google downloaded the sitemap within one second. Within an hour, 6 of the 10 URLs flipped from ‘unknown’ to ‘Discovered’.
End the island. One link from the main site to tech, sitewide in the footer, without nofollow. That was all.
Write titles and descriptions. On four pages of our main site, the page title turned out to have never been filled in. The CMS fell back to the page's name in the tree — which is why it said ‘opvallen | Admix’, lowercase and all. The same three pages also had no meta description. One page sat at position 2.6 with a 0.0% click-through rate. Position 2 to 3 should deliver 10 to 15%. Nothing was wrong with the ranking; a snippet had simply never been written.
Whether that was the cause, we'd know by 1 August. The answer came sooner — and our assumption turned out to be wrong. The breakdown per query showed that nearly all of that page's impressions were sitelinks on our own brand name: the extra lines under the main result, which almost nobody clicks. So there never was a snippet problem. The page simply didn't rank for any term beyond our brand name — a different diagnosis with a different treatment, which the page has since received: its own topic to be found on. That's how it works: hypothesis, one change, measure — and retract the conclusion when the numbers say so.
Of the ten pages that existed at the time, nine are now indexed. But our first explanation for that was wrong.
On 18 July, seven were in the index. The three missing ones were also the three shortest: 104, 153 and 212 words, against 260 to 576 for the pages that had been included. A clean pattern — too thin to keep. We wrote it down as the finding of the day.
One day later it no longer held. /marketplace/ (118 words), /nieuws/ (153) and the homepage (212) had all three been included, and their crawls dated from before the day we added content to them. They hadn't been too thin. They had needed more time.
What remained was one page, and it was more interesting than the pattern we thought we saw. /support/ had been in the sitemap since 17 July and for weeks was simply never visited. Not crawled and rejected — never assessed. That is not a content problem but a crawl problem: on a young host, Google spreads its attention, and a single page misses the boat. What finally pulled the page across the line was no trick, but a contextual link: one substantive paragraph on the main site pointing to it. Crawled and indexed within two days.
With that, the baton has been passed: the only page Google still hasn't fetched is this article. In the sitemap, internally linked, resubmitted — and yet Google picks its own moment per URL. Meanwhile, you're reading the proof that the rest of the approach works.
Why we're writing this down instead of polishing it away: this is exactly the mistake that makes so many SEO reports useless. Two data points on one day are not a trend, and the temptation to declare the first neat pattern the cause is strong. We fell for it ourselves, dashboard and all.
That doesn't mean thin pages are harmless — they just rarely give a reason to rank. But the blocker we thought we had found wasn't there.
More and more people put their question to ChatGPT, Perplexity, Copilot or Gemini instead of Google. That traffic behaves differently: fewer visitors, higher intent, and it shows up in no standard report.
We call that track GEO — Generative Engine Optimization. And it starts with the same question as SEO: do those systems fetch your site at all?
For us the answer initially appeared to be ‘no’. That turned out to be wrong. Our web server wasn't logging the field in which a visitor identifies itself at all, so a search for AI crawlers neatly returned zero hits across thousands of lines. Not because they weren't there, but because we weren't writing it down.
We enabled that field, and within five minutes the first AI crawler appeared in the logs. The first full day:
One observation stood out: ClaudeBot fetches our main site, but not the tech site. You only know that if you measure it. Without that log line, it's a feeling.
What we steer on for GEO:
To be honest: this field is young. Our baseline measurement of AI referrals was exactly zero, and we now measure monthly whether that moves. We'd rather sell a measured trend than a promise.
How such AI systems build their answers from a knowledge base — and why that doesn't necessarily require a vector database — is dissected in our article on RAG.
The observation above got a sequel. ClaudeBot ignored our tech site for five days — while the sitewide footer link was there the whole time. When we published one substantive paragraph on the main site containing a link to tech, ClaudeBot showed up in the logs the next day: 40 requests across 27 pages. Correlation is not proof, but the lesson is taking shape: a footer link is not enough for AI crawlers; a link in body copy is.
Two days later came the milestone we're really after: the first on-demand fetches. Not a crawler stocking up, but an AI assistant fetching a page live during a conversation with a user. Claude fetched a vacancy page; Perplexity and You.com followed the same day. Someone asked an AI a question about us — and the system came to get the answer at the source.
One caveat that keeps this field honest: those same logs also contained a visitor pretending to be ChatGPT that demonstrably wasn't. Any user agent can be forged. We only count an AI visit after verifying the sender; a raw count is an upper bound, not a measurement.
We've also put something in place ourselves: both sites now serve an llms.txt — a young file format that acts as a reading guide for AI systems: this is the site, these are the pages, this is what's on them. In our case the file is generated live from the same page list as the sitemap, so it never lags behind the site. Whether AI crawlers make structural use of it yet is unproven — we measure that too, and we'll report the outcome, whichever way it falls.
Everything above — the zero impressions, the 21 sitemap errors, the never-written title, the 0.0% click-through rate, the word-count pattern, the AI crawlers — comes from two sources: Google Search Console and your own site's analytics.
Without that access, SEO is guesswork. With it, it's diagnostics. The difference between ‘we recommend more content’ and ‘these three pages are too thin to be included, and here is the threshold where it flips for you’.
Two things we encounter in almost every engagement:
Your numbers are more incomplete than you think. Our own analytics show about 42% of what Search Console measures — the difference is visitors who don't consent to tracking. That's not a bug, but you need to know it before you interpret a dip.
Zero is usually a measurement error. Five times in one day we silently got back a zero that wasn't a finding but a misconfigured measurement. ‘Zero AI bots’ turned out to be an unlogged field. ‘Empty sitemap’ turned out to be an unfollowed redirect. First prove that you're measuring at all.
And watch who comes by: among our ‘Googlebot’ requests was a visitor asking for a password file. Anyone can pose as a search engine. A raw count is an upper bound, not a measurement.
We work in a fixed order, because skipping a step makes the next one pointless:
That is not a subscription to reports. It is a cycle with an outcome you can verify.
We're happy to take a look. With read access to your Search Console and your site analytics, we can often tell within a day where things are stuck for you — and whether that is a technical, a content or a findability problem.
Call us on +31 10 270 75 75 or email hallo@admix.nl. Already working with us? Then support works too.
SEO focuses on your position in the search results of Google and Bing. GEO — Generative Engine Optimization — focuses on whether AI systems such as ChatGPT, Perplexity and Copilot fetch your site and use your answer. The technical foundation largely overlaps: both start with a site that can be crawled. The difference is in the shape of your content and in how you measure it.
That depends on your starting position. On our own tech site, 6 of the 10 URLs flipped to ‘Discovered’ within an hour of submitting the sitemap. A day later 7 were in the index, the day after that 9. On a site Google already knows well it goes faster; on a brand-new domain without inbound links it can take weeks. So think in weeks, not days — and don't draw conclusions from the first measurement.
First check whether it has been fetched at all, because those are two different problems. If there is a crawl date, Google has seen the page and decided not to include it — usually because it resembles another page too much, gives too little reason of its own to exist, or has a canonical pointing elsewhere. If that date is empty, nothing has been assessed and it's a matter of crawl budget: Google simply hasn't come by yet. For us, the latter applied to exactly one page, while we initially thought it was about length.
Because it is the only source that tells you what Google really thinks of your site: which pages are included, which queries you appear for, what position you hold and how many people click through. Without that access we can only give you generic advice. With it, we can point out which page is the problem in your case. Read-only access is sufficient.
For Google's search results, FAQ markup no longer produces an enhanced display since 7 May 2026. Other types do — job postings, products and articles are still used extensively. And structured data remains the clearest way to tell a machine what's on your page, which is becoming more important for AI systems, not less.
No, and you should distrust anyone who promises that. AI systems don't give the same answer every time and publish no ranking. What we do instead is measure: do those systems fetch your site, does traffic come from them, and are you mentioned for a fixed set of questions we repeat monthly. That produces a trend instead of a promise.
llms.txt is a young file format: one overview on your site that tells AI systems in one go which pages exist and what they're about — like a sitemap does for search engines. There is no proof yet that AI crawlers use it structurally; it's cheap to serve, so we do, and we measure whether it gets fetched. One condition: have it grow with your site automatically, otherwise it becomes a second truth that goes stale.