The captures
The exact bytes we fetched — static HTML, rendered DOM, robots.txt, headers — stored content-addressed and never rewritten. We can show you what we saw, not a description of it.
If we have measured you wrongly, we want to know — and we can prove what we saw.
AgentProbe publishes quantitative claims about real websites. Some of those claims are adverse. A published measurement that cannot be disputed is not a measurement, it is an assertion, so this page says how to dispute one and what happens next.
Today that means the published figures of the current study — the 680-origin measurement at /findings/air-2026-08. The measurement of 26 August is no longer published on this site — its figures, its corrections and the evidence behind them stay in the repository, under reports/cp3b/ and reports/cp-pub/. No score is published yet, so no score can yet be disputed. This path is written to cover whatever is published, because a correction route that appears only once there is something to defend is not a route.
Every figure is reproducible from stored evidence.
The conditions
When the crawl ran, from which network and region, under which crawler identity, and which pages were sampled. A challenge decision often depends on where the request came from, so this is part of the finding.
The model
Which scoring version produced the number, and every weight and anchor in it, published in scoring/v11.yaml.
What to send, and what happens to it.
Write to bot@agent-probe.com with the origin and what you believe is wrong. You do not need to prove anything — naming the figure is enough to start.
01
Acknowledged within two working days
The same commitment the crawler page makes, because it is the same inbox and the same person. If a correction takes longer to resolve than that, you get the interim state rather than silence.
02
We reproduce it from the stored capture
Most disputes resolve here, and often in your favour without a re-crawl: the question is usually what the crawler was served, and we still have it.
03
If the dispute is about drift, we re-audit on the same pages
A re-audit is pinned by default — it re-crawls the original sample rather than drawing a fresh one, so the comparison is on the same documents. An unpinned re-audit is labelled a new measurement and never presented as a delta.
04
If we were wrong, the figure is corrected or withdrawn
Published figures are regenerated from the corrected data. Where a correction changes something already published, the change is dated and the previous value is stated rather than quietly replaced. The study already carries its own corrections, each stating what it changed and what it changed it to.
You can also just ask us to stop.
Our crawler identifies itself honestly and respects robots.txt, so blocking AgentProbe there is sufficient and needs nothing from us. If you would rather be on a permanent exclusion list regardless of what your robots.txt says later, write to the same address and you will be. See the crawler page for what it fetches and how often.
Blocking us removes you from the population; it is not recorded as a finding against you.
A refusal is a finding, and it is not a criticism
An origin that challenges or blocks the crawler stays in the population, because refusing automated readers is itself the thing being measured. It is frequently a deliberate and correct decision — the study reports what an AI system encounters, not whether you were right to make it encounter that.
A challenge is a fact about an encounter, not about you
Bot management is not deterministic and its decisions depend partly on the address space a request came from. A challenged origin has its comparable score withheld rather than lowered, precisely because we cannot tell the difference between a site that is hard to read and one that declined to be read by us that day.