What our logs say about the robots that read us
In short: we counted, robot by robot, who really comes to read the tower. Over thirty days: 2,486 requests claimed to be an artificial intelligence robot, and 758 came from an address that was not theirs. Nearly one liar in three.
1. The question we had never asked
We knew robots came. “ChatGPT reads us”, “Claude reads us” — we said it as a given, without ever checking line by line. So we went to the server log: the file that records, second by second, who knocked at the door, with what identity card.
2. Nearly one robot in three lies
First shock: a visitor presenting itself as “GPTBot” or “OAI-SearchBot” is not necessarily one. Anyone can write that name in a request. The only way to know is to check the address the request comes from against the list the real provider published. Of 2,486 requests claiming to be an “AI robot”, 758 came from elsewhere.
3. Less than one page in four exists for AI
Of the 313 pages in the sitemap, 72 were read by a real AI robot in thirty days. The other 241 simply don't exist for an artificial intelligence. Not because they are bad. Because nobody, human or machine, has come looking for them yet.
4. Two robots that don't look alike
A training robot (ClaudeBot, GPTBot) comes in waves, to feed a future model. A real-time search robot (OAI-SearchBot) fetches one precise page, right now, because a real person just asked an AI a real question — and that page served as the answer. That second kind matters most: each time it comes, someone has just cited us.
5. What we show, what we keep
We show here the result of the measurement: the numbers, the proof that a robot can lie about its name. We don't show the exact mechanism used to verify an identity — it stays a work tool, not a blog post.
We no longer guess who reads us. We measure it, robot by robot, and fix what the measurement shows.