What 90 days of request observations show on this website

What the observer actually recorded

The three classes sum to the same 49,152-row denominator. Ordinary-browser is a classifier result, not proof that a person was present. Likewise, an unmatched automation record has no established company behind it. Keep that uncertainty when comparing totals.

Which familiar names appeared?

These selected registry labels describe claims inferred from User-Agent patterns. The stored registry was version 2026-07-11 and the classifier was 2026-07-11.1 throughout the snapshot. A label can be copied by another client, and a later registry could classify an old string differently.

Another 15,923 self-identified-bot events had no named agent match. Leaving those unmatched is more informative than assigning them to a company from a weak hint. The full grouped table is in the snapshot bundle.

Ninety elapsed days do not prove ninety days of complete capture

The fixed interval intersects 91 UTC dates because it begins and ends partway through a day. At least one event appears on every intersecting date. That checks for whole missing date bins; it cannot rule out shorter outages, dropped deliveries or requests excluded before ingestion.

The middleware selects public GET and HEAD documents plus certain machine-readable resources. It excludes APIs, health checks, most static assets and other methods. Its delivery is best effort and fails open, so a logging problem should not interrupt a visitor. The saved rows are the successful observation deliveries, not an external traffic census.

The window can include the owner’s browsing, synthetic checks and diagnostic clients. No complete organic-versus-test label was available in this extraction. Dividing these counts by days produces an accepted-event average, not a provider’s true crawl rate or a measure of audience size.

What decisions can these records support?

They identify names and routes worth investigating. A surge in a claimed agent can lead to an edge-log check; an unmatched class can lead to classifier review. Before changing a provider-specific rule, follow the identity-verification method and inspect what the request actually received.

They do not establish whether a provider respected robots.txt. That requires the delivered policy at the time, the applicable token and path, a defensible client identity, and a controlled interpretation of caching and user-triggered fetches. The robots-compliance evidence guide explains the missing joins.

Nor does requesting a document establish reading, training or answer citation. The narrower llms.txt study shows how that distinction changes even a simple-sounding question.

Reproduce the arithmetic without exporting private request records

The download includes daily class totals, claimed-agent aggregates, llms path/method counts, version counts, the aggregate SQL and a method note. Sum the daily table by observed class to reproduce the headline. Sum the claimed-agent groups to reconcile the 30,348 non-browser-classified events.

The public projection omits raw events, full user agents, IP-related fields, geography, query names, referrers and deployment identifiers. A later database query can change as ingestion and retention continue; use the frozen files for this edition’s arithmetic. No other website is included.