Measure AI visibility: one reading does not hold

What do we look at on our own pages?

Three things nobody looks at: what survives extraction, who actually comes to read us, and which of our pages surface in the answers.

1. What survives extraction

An engine does not read a page, it reads what an extractor pulls out of it. We fetch the page with an AI crawler user agent, exactly what it receives, then run it through the three reference extractors: trafilatura, jusText and Resiliparse. The three pages below belong to software vendors, two in HR software and one in workforce scheduling.

On HR software vendor B, trafilatura keeps only 23 % of the served text. Four fifths of what the page says never reaches the engine. And the three extractors do not agree with each other: on that same page one keeps 23 %, another 40 %. Content can therefore reach one engine and vanish from another.

2. Who actually comes to read us

A crawler's declared user agent can be forged in one line. So we check every request against the IP ranges the vendors publish. Here is the result on our own servers, over seven days.

Out of 1,734 requests declaring an agent whose vendor publishes its addresses, 1,615 came from elsewhere, or 93 %. The ones presenting themselves as Perplexity were Google Cloud machines and a hosting provider, while Perplexity publishes only eight addresses.

Not all these crawlers want the same thing

Counting them together mixes two subjects with different consequences. Some collect training material, others build a search index, others read a page at the moment a user asks a question.

The consequence lands straight in robots.txt. Blocking a training collector costs no citations. Blocking an index crawler or a user-triggered reader removes the site from answers. Many sites block the second kind believing they are blocking the first.

One crawler in twenty hits a dead page

Over the same seven days we counted the share of requests that receive a 404 error. The result varies enormously from one crawler to the next.

The public reference measurement, published in late 2024, reported 34.8 % for OpenAI crawlers against 8.2 % for Googlebot. On our servers it is 4.9 % against 1.5 %. The order is the same, the scale is six times smaller, and that is exactly why a public measurement does not replace your own.

A dead URL is a lost source. The redirect plan, treated everywhere as a technical chore, is therefore a visibility matter.

3. Which pages get picked up, and which never

The question arises page by page, and the answer is surprising. A page with heavy search exposure may never be picked up, while a confidential one may be picked up one time in five.

The Observatory is our most picked-up page, at 20.6 % of its impressions, for 34 impressions in total. The article on the LinkedIn algorithm has 1,043 and is picked up only 1.9 % of the time. Search volume and engine pick-up do not follow each other. A page can also be read without ever being cited, like our consulting page, present in search and absent from every generative answer.

White paper, counting bots →
White paper, the recoverable message →

5. What do we fix when the measurement is bad?

The two grounds above produce findings. Our audit tool turns them into work. It scores each page across four families and lists, rule by rule, what blocks: a lede too short, a title that does not carry the subject, a FAQ missing from the markup, a paragraph that refers to another without naming it.

On the page you are reading, the audit of 3 September 2026 puts the technical and SEO families above the advised threshold, and GEO below it. We leave that visible rather than remove proof to gain points: the tables add sections without adding questions, which the rule penalises. We preferred the proof to the points.

The four scores change with every edit to the page. Publishing them frozen in a table would mean publishing a stale measurement from the next revision onwards, which is exactly what we criticise in others.

What correction changes, at the scale of a market

These findings are not specific to us. The Message-Market Fit Observatory audits the sites of French startups after fundraising, on a grid of clarity and engine readability.

Three startups in four fall below the threshold, and the money raised changes nothing. The last figure is the most useful: message clarity cannot be bought with a funding round, it has to be worked.

Fixing and verifying are two different jobs

Confusing them wastes time. The first instruments say what to change on a page, today. The second say whether the change produced an effect, weeks later.

The cost makes the requirement affordable. With the provider we use, one question costs four thousandths of a dollar, so 0.20 dollar for fifty. We publish that figure because a measurement at that price has no reason to be replaced by an assumption.

The Google AI Overviews barometer, method and editions →
ChatGPT visibility, what the engine sees of your site →
876 impostors, what your AI traffic really hides →

7. What would the dashboard look like?

Like this. Everything below comes from the measurements described above, on fast-growth.fr, between 1 and 2 September 2026, each block from a different instrument.

Three levels, in the order a director reads them. At the top, what Google answers on our market and whom it gives the floor to instead of us. In the middle, which of our pages it picks up and which it ignores. Only at the bottom, how it happens: the crawlers that visit, the composition of the sources and what the extractors keep.

The order is not decorative. A dashboard that opens with crawler names makes someone read plumbing when they are trying to find out whether their market knows them.