About

Why this exists, and what it refuses to become.

The gap this fills

A large and growing share of web traffic is not people. It is crawlers assembling training corpora, agents fetching pages to answer a question someone asked a minute ago, and automation of every other description. Site owners are told to optimise for this traffic, and are sold measurements of it that are produced by asking an AI system what it thinks — which is the least reliable instrument available for the question.

Meanwhile the basic facts are not publicly established. Which crawlers honour the rules they are given. Which execute JavaScript. How long it takes for a newly published sentence to become something a model will repeat. These are answerable by observation, and largely unanswered in public.

What this is

One domain, instrumented deliberately, publishing what it sees. The site is both the instrument and the subject: every claim made here was measured here, which keeps the claims small and keeps them checkable. There is no customer data involved and no privileged access to anything.

What this is not

• Not a security product. Nothing here blocks, challenges or scores a visitor. Observation is the entire response.
• Not an AI visibility tool. Those work by asking models about you. This observatory treats that as inadmissible, which rules out the whole category.
• Not a threat feed. No reputation lists, no naming of individual actors, no fear marketing. Automated traffic is a population to be described, not an enemy to be sold protection from.
• Not finished. Observation began 2026-07-25. Most of what this site intends to know, it does not yet know.

How to read what is published

Every figure on the lab page traces to specific recorded requests, and the methodology lists the ways each one could be misleading. Where the sample is too small to support a conclusion, the page says so rather than rounding up to a finding. The constitution states the rules the whole thing is bound by.

Corrections

There is no contact route here yet, and saying so is better than publishing an address that does not receive mail. One will be added when it works.

What a correction needs, whenever there is somewhere to send it: the page, the figure, and the record you believe contradicts it. A correction that lands gets published as a correction rather than folded into an edit — the findings already carry two withdrawn conclusions and six rejected ones, with the reason kept beside each.

There is no form on this site, no newsletter, and no list. A page that collects addresses would be doing something to its visitors, and everything here is built to do nothing but watch.