Ready for Google, invisible to AI: what 200 scanned websites reveal

Two years ago “being findable” was one question: are you in Google. Today it is two, because a growing share of your audience asks ChatGPT, Perplexity or the AI Overview above the search results. Those two questions are rarely measured together. We did measure them together, across 200 randomly drawn Belgian websites, and one figure sums up the outcome: of those 200, exactly one reaches a good score on AI visibility.

How wide is the gap between SEO and AI visibility?

Almost thirty points. Across 200 measured websites the average SEO score is 69.1 and the average GEO score is 39.5, on the same scale of one hundred. Classic findability is therefore reasonably in order, while findability for AI answers lags far behind. That gap of 29.6 points is the core of this research.

The distribution makes it sharper than the average does. A third of the sites reach 80 or more on SEO. On GEO, one site out of 200 manages that.

Put differently: almost six in ten sites sit below forty on GEO, while only just over one in ten sit there on SEO. The median confirms the picture, with 70 on SEO against 35 on GEO.

Why does almost every site score lower on GEO than on SEO?

Because the two disciplines are of a different age. Of the 200 sites, 195 score lower on GEO than on SEO, which is 97.5 percent. Five sites do the opposite. Such a one-sided outcome does not point to chance but to a structural difference in what gets delivered by default.

Classic SEO has been baked into every common system for twenty years. A modern site gets its title, viewport and indexability automatically, or through a plug-in someone installed once. Our measurement confirms it: 2 percent fail on indexability, 4 percent on the viewport tag and 10 percent on the lang attribute. That is the legacy of two decades of tooling.

The signals AI systems lean on are considerably younger and sit in no standard installation. Structured data about who you are, what you stand for and who is behind your company has to be added deliberately. There is no checkbox that switches it on. That is exactly where the score falls away, and it explains why the gap points so systematically in one direction.

Does the gap narrow for sites that do have their SEO in order?

No, it widens, and that is the most striking outcome of this measurement. The better a site scores on classic SEO, the wider the gulf with its AI visibility.

Read that last row again. The 29 sites that exceed 90 on classic SEO, the top of the class, stall at 50.3 on AI visibility. Their GEO score is barely higher than that of the group with an SEO below 60, while their SEO score is more than double.

Performing well on classic SEO does not, in other words, bring AI visibility along with it. That is counterintuitive, because it concerns the same pages and often the same people working on them. The explanation lies in the nature of the work: whoever takes SEO seriously optimises titles, speed, internal links and content. None of those four touches the question of whether a language model understands which entity sits behind the site.

Which AI signals are missing most often?

The five most frequently missed signals are all five GEO signals. Only in sixth place does a classic SEO check appear. The column “entirely absent” counts the sites where the signal is not there at all; “not in order” adds the sites where it is incomplete.

Three rows carry a dash in the first column. Those signals cannot hard-fail in our scan: they are either present or absent, and absence produces a warning rather than an error. In substance, 99 percent there simply means: on 197 of the 200 sites there is not a single question with an answer underneath it.

The most striking row is the seventh. On 119 of the 200 sites there is no structured data whatsoever. Not a single block of JSON-LD, anywhere. That is not a matter of an incompletely filled field; that is a site that presents itself in no machine-readable way at all.

What does a missing Organization schema mean in practice?

That an AI system has to guess who you are. The Organization schema is the block of structured data that literally states: this company is called this, sits at this address, has this logo and can be found on these profiles. Without that block a model can only derive your brand name from running text, and then your company name competes with every other occurrence of that same word on the internet.

The figures are at their sharpest here in the whole study. Of the 200 sites, 153 have no usable Organization node at all, well over three quarters. Another 38 do have one, but incomplete: name and URL are there, and then the logo, the contact details or the sameAs field is missing. Exactly nine sites out of 200 have a complete Organization schema. That is four and a half percent.

That sameAs field is the most underrated part of it. It is the list of your other official profiles, from LinkedIn to your Google business profile, and for an AI system it is the way to confirm that the company on your site is the same company as the one in its training data. Leaving out that field means leaving out the confirmation. How to build such a node correctly is covered in our guide to implementing JSON-LD.

Why do nine in ten sites miss a Person and a FAQ signal?

Because both only became useful when AI answers arrived, and the classic SEO checklist never mentioned them. On 196 of the 200 sites a Person schema is absent. On 197 sites both FAQ markup and a visible question-and-answer section are missing. On 158 sites there are no question headings in the text.

The Person schema links a real name to your company: the founder, the author, the expert. For a language model trying to assess whether a source is trustworthy, a named human is a stronger signal than an anonymous company page. Four sites out of 200 have that.

The FAQ finding needs a caveat you rarely read elsewhere. Google has shown almost no FAQ rich results since 2023 and removed the documentation about them entirely in June 2026. For classic SEO, FAQ markup has therefore become close to worthless. For AI visibility it has not, because a question with a short and complete answer underneath is exactly the format in which a language model cites. Anyone who removed FAQs because the rich results disappeared has unknowingly given up a citation opportunity.

How many sites have the basics fully in order?

None. Take the four core signals together, so structured data present, a complete Organization schema, a WebSite schema and a Person schema, and zero of the 200 sites pass on all four. Five sites manage three of them.

At the other end of the distribution sit 119 sites that pass on none of those four signals, sixty percent of the sample. For an AI system such a site is in practice a sheet of text without an identity card: readable, but not attributable to a recognisable organisation.

That zero is the figure that stays with you. These are not stragglers or forgotten corners of the web: these are two hundred Belgian sites with measurable traffic, and not one of them has the four basic signals an AI system needs to recognise you with certainty.

So does nobody reach the full score?

Someone does, but we had to look outside the sample for it. Our own site scores 100 out of 100: all ten SEO checks and all ten GEO checks pass, measured with exactly the same scanner on the same day.

Straight to the caveat you would rightly raise: clickforest.com was not among the 200. The domain was not drawn and does not even appear in the source list we drew from, so it influenced the figures above in no way whatsoever. We measured it separately, precisely because the question “and what about you then?” deserves to be asked of an article like this.

And the second caveat, which matters more: it is our own yardstick. A perfect score on your own instrument is by definition the easiest number there is. So what follows is not the figure but the method, so you can check it with whichever measuring tool you prefer.

What is actually in place is exactly the list the 200 measured sites are missing. Site-wide structured data, with a complete Organization schema including logo, contact details and a sameAs list to the real profiles. A WebSite schema, and a Person schema for the author, so that a named human stands behind the text. Beyond that, the subheadings are real questions with an answer underneath that reads on its own, there is a FAQ block on every important page, and the company name simply appears in the running text instead of “we” everywhere. On top of the classic basics most sites already have.

None of those things is expensive or complicated. They are merely rarely done, and that is exactly what this research shows.

How was this research carried out, and what does it not say?

The sample is a random draw from the Belgian .be domains in the Tranco list, a research ranking of websites, in the version of 17 August 2026. The 250 largest .be sites were skipped, because those are search engines, banks and media, not everyday websites. From the rest a draw was made with a fixed seed, so that anyone who wants to check the draw gets exactly the same 244 domains.

Of those 244, 200 were measured on 18 August 2026, all on the same day. Every site was fetched once with ordinary browser settings and immediately tested against 20 checks: ten on classic SEO and ten on AI visibility, each with its own weight. One page per site was measured, usually the homepage.

Why 44 sites were not counted, and why that figure appears here. Twenty-four were unreachable, thirteen refuse automated requests, and five redirect to a non-Belgian domain. Those thirteen blockers deserve attention: sites with such a security layer in front are often precisely the professionally managed ones. Silently leaving them out would tilt the sample towards more poorly maintained sites and flatter this article’s conclusion. Hence the figure appears here rather than in a footnote.

Now the limits, because they determine what you may do with these figures.

The sample consists of websites with measurable traffic. The completely invisible site nobody ever clicks is not in it. That makes the outcome stricter rather than milder: if even sites with an audience miss the AI signals, the finding is stronger than if we had only measured forgotten corners.

There is no filter on the type of organisation. Webshops, SMEs, media and public services are all in it. An earlier attempt to select on that failed: the automated classification was wrong in both directions, and deciding for yourself who gets into your own research is exactly what makes a sample unreliable. This is therefore a cross-section of the Belgian web, not a sector study.

The scan measures one page per site. A site may well have a FAQ or a Person schema further along. For the schema signals that matters little, because those belong site-wide, but for the content signals it is a real limitation.

And the scores come from our own model. Which ten signals count towards the GEO score and how heavily they weigh is our assessment of what AI systems use, based on the documentation of schema.org and Google plus the research into generative engine optimization. There is no official GEO standard to test against.

Finally the margin. At 200 measurements the 95 percent confidence interval around a proportion of fifty sits at roughly 7 points, and around a proportion of ten at roughly 4 points. The large outcomes in this article, such as that 97.5 percent and that zero out of two hundred, sit far outside that margin.

What can you do with this in the coming week?

Four things, in this order, because that runs from largest effect to smallest effort. They follow directly from the tables above.

Start with structured data, full stop. Sixty percent of the measured sites have not a single line of it. Put a complete Organization schema there with name, URL, logo, contact details and a sameAs list to your real profiles. That is the signal 96 percent of sites fall short on, and it answers the most basic question an AI system asks about you. Then add a WebSite schema and a Person schema for the founder or the author, which is missing on 98 percent of sites.

Next, put your most important customer questions as real question headings on your service pages, with an answer of three or four sentences underneath that reads on its own. That is no longer an SEO trick, because the rich results are gone; it is the format in which you get cited. More on that approach is in our GEO strategies and in the glossary for the terminology.

And finally, simply name your own company name in the running text, instead of writing “we” everywhere. On 56 percent of sites the brand name is insufficiently anchored, and a model that does not encounter your name in the text has nothing to quote.

Conclusion: what this research actually shows

That most websites are not badly built, but built for the previous question. They neatly answer “am I in Google” and leave “does an AI system understand who I am” unanswered, and at 29.6 points that difference is larger than most owners suspect. It moreover sits not in the content or in the design, but in a handful of lines of structured data that nobody ships by default.

What struck me most while working through this: the sites that score best on classic SEO have the widest gap, and that relationship is no coincidence but rises steadily. Taking your findability seriously therefore offers no automatic protection, and should in fact be the first reason to look. If you want to know where your own site stands, the GEO audit covers the same ground as this research in full. How to tackle the findings afterwards is in our approach to GEO. And how your brand ends up in AI answers is covered in getting recommended by ChatGPT.