Every request reaching the study's server is placed in exactly one of three categories. Only the third supports any conclusion about the client.
| Category | Meaning | Counts against the crawler? |
|---|---|---|
| First sight | This client has not requested this page before. It held no copy, so a full download was its only option. | No |
| Return, unstamped | The client returned, but the previous response it received carried no version fingerprint. It had nothing to revalidate against. | No — this is a fault of the server, not the client |
| Judgeable | The client returned to a page for which it had previously been given a fingerprint. It could have asked; the question is whether it did. | Yes |
Within the judgeable category, every visit is either an ask (the client sent a revalidation request) or a blind re-download. These two always sum to the judgeable total; no visit is counted twice, and none is omitted.
A measurement that simply divided revalidation requests by total requests would be biased by construction. A crawler discovering a hundred new pages would score as fully non-revalidating while having done nothing wrong, because a first visit cannot carry a revalidation request. Excluding those visits is not a courtesy to the crawler; it is the only way the ratio means anything.
Where a crawler has no judgeable visits at all, the study reports that no conclusion is available rather than reporting zero. Zero and "no evidence" are different findings and are presented differently.
Googlebot is used as a control, being documented to issue revalidation requests. If the control appears not to revalidate, the fault is assumed to lie with the study's own server configuration until shown otherwise.
Part of an open study of AI crawler cache behaviour. Study reference RCP-2026-08. These pages are served by the study's own logging server; requests to them are recorded for the purpose described in the method. Content is stable by design and is not edited during a measurement run.