My structured data said exactly who I was. My visible text said nothing.
For a language model, that is no minor distinction. It reads what is written, not what you mean — and what only exists under the bonnet does not exist for a model that reads. I made that discovery when I put my own site through the audit process I built myself. But let me start at the beginning.
More and more searches no longer start at a search bar but with a question to ChatGPT, Gemini, Claude or Perplexity. The answer is not a list of ten links, but one composed narrative with a handful of sources. For anyone who wants to be found online, the question shifts accordingly. No longer “do I rank high in the results”, but: am I used in the answer?
That is an uncomfortable question, because no dashboard tells you why you were skipped. A page that scores perfectly in classic SEO can be completely invisible in an AI answer. I wanted to know why — and the most honest way to find out was to audit my own site. Publicly. Gaps included.
This is what came out of it.
I. An AI answer passes through four gates
Before your page ends up in a generated answer, it has to pass through four gates — that is the model I use to look at AI visibility, and the one the audit later in this piece is built on. Each gate is a question the system must answer with “yes” before the next one counts.
The first is access: can the system retrieve and process your page at all? If a crawler is blocked or your content only appears after scripts run, the rest is irrelevant.
The second is retrieval: when a concrete question is asked, is the right passage on your page found? A page can be reachable and still never surface, because it does not connect with how people phrase their questions.
The third is extraction: is there a self-contained, quotable answer? One sentence that stands on its own, without the paragraph around it. If a model has to reconstruct an answer from your text rather than lift it out, it will rather do so with a competitor who offers it ready-made.
The fourth is selection: does your answer beat the alternatives? Specificity and a defensible claim decide which source the system picks when the answer is already nearly complete.
The key insight: failing at an early gate makes everything after it pointless. A closed access gate makes your quotability irrelevant. There is no point polishing the finest details while a fundamental gate is still jammed.
II. My own homepage: 63 out of 100
No disaster, no cause for celebration. A page with a solid classic foundation, but with a few fundamental gaps that hit those gates precisely.
Three things stood out.
My page nowhere said in plain words what the platform is and who it is for. That information did sit in the structured data under the bonnet — but what a page merely implies, a model might guess. And guessing is not what it cites. The model reads what is written, not what you mean.
There was no self-contained answer up front. The information sat woven into narrative paragraphs — pleasant for a human reader, difficult for a system looking for one quotable sentence.
And there was no FAQ answering the follow-up questions people actually ask: what is this exactly, how does it differ from that, how do I measure it. Questions an AI happily cites a ready-made answer to — if it is there.
III. What I changed — and what I deliberately ignored
I executed the shortest list that opened the weakest gates. No major rebuild, one focused afternoon of work.
I placed one factual sentence at the top that says in plain language what Groundbase is and who it is for. I gave the page an author — my name, visible in the text and not only in the metadata, because who stands behind a claim carries weight. And I added an FAQ: five questions, each answered in the first sentence, with the context following after.
After the changes: 63 became 72. Is that a good score? Wrong question. What matters is that the nature of the findings shifted: the audit no longer spoke of missing building blocks, but of polish. From fundamental to refinement. That is a category change, not a points gain.
The most honest part of this exercise, however, is the recommendations I did not follow. The audit suggested, among other things, a services section, and a rewritten, more literal main heading. Both defensible from the system’s point of view, both deliberately ignored — the first because my positioning is that of a knowledge platform, not a service provider, the second because my heading carries brand voice I was not willing to trade for a search engine.
Note, August 2026. This piece says I deliberately chose not to follow the recommendation to add a services section, because my positioning was that of a knowledge platform, not a service provider. That has since changed: Groundbase now openly offers services. The audit finding stands — the point is not that a services section is always required, but that what you visibly put on your page steers what a model says about you. Back then, absence was the right answer to who I was. Today, presence is.
That is precisely the point. An audit gives you the shortest route, not a law. You decide which gates you want to open. A tool that forces you towards a perfect score often forces you towards a worse page.
IV. Validated on paper, invisible in practice
The first gap from the audit deserves its own section, because it is the finding that stuck.
On paper, everything was right. My schema markup neatly described who I was, what Groundbase is and who is behind it. The validators showed green. Every technical checklist would approve this page — and that is exactly why the gap stayed invisible for so long.
But the four gates do not run on markup alone. Extraction looks for a sentence a human could read too: self-contained, quotable, in the visible text. My name and positioning existed in the code and not on the page — and so they did not count where it matters. Structured data supports what is written. It does not replace what is missing.
That is the most dangerous kind of problem: invisible, because every tool tells you it is fine. What you leave the reader to conclude, a language model does not conclude. You only find this gap by reading your page the way a model reads it — not the way a validator ticks it off.
V. Three questions that take you further than most tools
You do not need to build a tool to benefit from this. Three questions take you a long way.
Do your most important pages say in plain words what you are and who you are for — not implied, but stated? Does the answer to the core questions sit at the front of the text, readable on its own, rather than buried in a paragraph? And can an AI system actually retrieve your page — not according to your configuration, but tested in practice?
Those three take you through the first gates. The rest is refinement.
The audit in this piece comes from Answera, the tool I built to measure how AI systems read a page. The four gates in this piece are its four layers — the tool is the methodology, made executable. It is currently available by invitation. Questions about AI visibility or about the audit: hello@groundbase.be.
Because that is the lesson that lingers, beyond schemas and validators too: what is right on paper is not yet what gets read. And a page that thinks it is ready for AI only truly is once an AI confirms it.