Methodology

What counts as evidence on this site, how it is collected, and where the method is weak.

Admissible evidence

One thing only: a request this server observed and recorded. Method, headers, timing, status, bytes. Records are written once and never modified, and they carry no verdicts — a stored judgement would overwrite the evidence needed to check it.

Inadmissible evidence

Model self-report. We do not prompt a language model about this site and record its answer as data. A model describing its own knowledge is the system under test testifying about itself; it can assert familiarity with pages it never read and deny ones it did. This exclusion rules out the common approach to AI visibility measurement, which is the point rather than an oversight.

Inference presented as observation. That a client "is a bot" is a conclusion. That a client sent a given User-Agent and fetched given paths at given times is an observation. Only the second is stored.

Interpretation is separate and versioned

Signals derived from observations are written to their own table, tagged with the version of the interpreter that produced them. Improving the interpreter means deleting the prior version's output and recomputing over the same untouched observations. A conclusion that cannot be re-derived from the record does not get published.

Known limits

• User-Agent strings are unverifiable. Anyone can send any string. Counts by declared agent are counts of claims, and some fraction of traffic claiming to be a given crawler is not.
• Robots compliance is measured imprecisely. A fetch of a disallowed path proves the fetch happened, not that the client read robots.txt and chose to ignore it. Distinguishing those requires correlating a robots.txt fetch with the same client beforehand, which is only reliable when the client is consistent.
• One site is one sample. Everything here describes how AI clients treat this specific domain — new, small, and about a niche subject. Generalising to the web at large is not supported by this dataset.
• Absence of a marker proves little. A coined string not appearing in a model's output may mean the page was never ingested, or that it was ingested and not surfaced. Appearance is strong evidence; non-appearance is weak evidence.
• Client IPs arrive via a proxy. The site is served through Cloudflare, so the connecting address is taken from a proxy-supplied header. Both that header and the raw socket address are stored, unmodified.

Privacy

No cookies, no fingerprinting scripts, no cross-site tracking, no accounts. The unit of observation is a request. Published tables report user agents, paths, and counts — never anything that identifies a person.

Reproducing this

The instrument is a small server that records requests and a set of pages designed to make specific behaviours measurable: disallowed paths for compliance, a script-rendered page for JavaScript execution, several formats of identical content for format preference, and coined markers for ingestion. None of it requires special access. Any site owner can build the same thing.