Week 2026-W30
2026-07-20 00:00:00 UTC to 2026-07-26 23:59:59 UTC. Every figure on this page is computed from the record when the page is requested, not written down afterwards.
There is no previous week to compare against.
The largest single address contributed 90 of those requests — 15.7%. Anyone can add to this count; nothing here refuses a client, because refusing traffic would change what the instrument can measure. So the concentration is published beside the total rather than prevented.
Declared identities
| Declared an AI crawler identity | Requests |
|---|---|
| Corroborated by the vendor's published range | 7 |
| A different range belonging to the same vendor | 0 |
| Contradicted — vendor publishes a list, this address is not on it | 30 |
| Uncheckable — no published list exists | 23 |
| Total | 60 |
Checked against the vendor snapshot captured 2026-07-26, never a live fetch, so this table reproduces. Only contradicted is evidence against a client, and uncheckable is a gap in a vendor's publishing rather than anything about the client. The running totals carry the same split.
Behaviour
| Measurement | This week |
|---|---|
| Fetches of paths disallowed in robots.txt | 0 |
| Requests carrying a conditional header | 1 |
Both figures are small and both are honest about why. The disallowed paths serve ordinary content and are listed so that compliance is measurable rather than assumed. Conditional requests can only be counted on the formats where the validator survives the CDN — which is not all of them.
Findings decided
7 published:
A new domain with no inbound links received its first automated probe in under two minutes
agentshieldaidefense.com began serving at approximately 10:43 UTC on 25 July 2026 with no inbound links, no search index presence and no sitemap submitted anywhere. The first unsolicited request arrived at 10:44:54 UTC — roughly two minutes later — probing for a WordPress installation path. Four further independent sources followed within the hour.
An assistant given a direct URL substituted a search, and reported a competitor instead
Asked to open a specific URL on this site, a commercial AI assistant did not request the page. No corresponding request reached the server. It performed a search instead, found an unrelated company with a similar name, and reported that the requested page was unreachable or unindexed. A second attempt, explicitly instructing it not to search, produced a real fetch.
Anatomy of a user-triggered fetch: one page, no robots.txt, no JavaScript
The first verified AI agent visit to this site fetched exactly one page, did not request robots.txt beforehand, executed no JavaScript, and did not follow a single link. It declared ChatGPT-User in its user agent and connected from cloud infrastructure in Switzerland. Ingestion was confirmed independently by a coined string the client could not have guessed.
One user agent, 206 addresses, 58 paths — a retrieval spread thin enough to look like nothing
A single user agent string arrived from 206 distinct addresses in 10 countries over 73 hours. 145 of those addresses sent exactly one request. Between them they fetched 58 distinct paths.
We asked Claude-User to read this page, and it did
A client declaring Claude-User made 4 requests across 4 distinct paths on 2026-07-25. Every one of them arrived inside a window in which we had asked this vendor to read a page here, so this is the result of a trial rather than an unprompted visit.
Machines are fetching this site through a second hostname of ours that we never published
Our server answers to two names. Everything this site publishes — robots.txt, the sitemap, llms.txt, the feed, every canonical tag — uses agentshieldaidefense.com and never www.agentshieldaidefense.com. Even so, 159 fetches arrived asking for www.agentshieldaidefense.com, 91 of them from machines that do not share our own browser language settings. They did not learn that name from us.
One address requested 87 distinct paths within 6 seconds
A single connecting address took 87 distinct paths in 6 seconds, a rate no interactive session produces.
7 withdrawn or rejected. The subject and the reason for each are listed on the findings page.
Rejections are counted here and never named, because publishing an unverified statement about a client in order to prove it was rejected repeats the offence. The count is the part that matters: a review step that has never rejected anything has not been shown to be one.
Markers
47 new markers published this week, 75 live in total. 0 have ever been observed in a language model's output.
That last number is the one this site exists to move, and it has been zero since the first day. It is printed every week whether or not it changes, because a counter that only appears when it is interesting is a counter nobody can trust. What a marker is.
Other weeks
- 2026-W31
- 2026-W30 — this week
Each week keeps its own address permanently, so a figure quoted from here can be checked against the week it was quoted from. Reports are recomputed on every request, which means an older week's page may change if the record behind it is corrected — the figures are a view of the record, never a snapshot taken away from it.