AgentProbe asks your origin for the documents written for AI agents, the way an agent would ask, and reports what came back. Not what your CMS believes it published — what was delivered.
Three policy endpoints, two client identities, six requests. It reports the statuses your origin returned and whether your edge treats a declared agent differently from a browser. It fetches no content pages, runs no browser, and will never report a score — the check page says exactly what it asks for.
A page can be excellent for one and poor for the other, and the fixes have almost nothing in common. Keeping them apart is the whole method.
Whether an engine can obtain and use what is on the page. Can it be fetched? Is it there before JavaScript runs? Does an extracted passage still make sense alone?
This is what search and answer engines do when they cite you.
Whether an agent can perceive, navigate and act on the interface. Can it identify the controls? Does it know what they do? Did the task complete?
Measured and reported, but carrying no weight — nothing yet shows that changing these changes what an agent can do.
A study asks one question of a fixed corpus in a fixed window, publishes every figure with its denominator, and stays citable afterwards. There is no continuous benchmark behind this site, and nothing here will imply one until there is.
The current study measures what 680 origins served a declared AI agent and a browser within the same seconds, from two networks. It reports the hypothesis it failed to confirm first, because that is the one that failed.
Naming these here is what makes everything above them believable.
The scoring model is specified in full on /method — attributes, dimensions, one gate, a 0–100 result. Nothing on this site computes one yet, and no origin has been scored or ranked.
It asks three policy documents from one machine in eu-west-1, twice — once as a declared agent, once as a browser. The study behind this site used two networks and five client identities; a single check cannot see what varies between them, and a result is not evidence that an origin behaves this way for everyone.
A submitted name is reduced to the domain somebody registered, so blog.example.com is answered as example.com — unless the apex has no address, in which case the name you typed is used. robots.txt is per host, so a site serving different rules on a subdomain is not measured there.
Everything published here comes from a single measurement window in August 2026. There is no second run, so “is this getting worse?” has no answer here and this page will not imply one.
If a published figure here is wrong, the corrections process says how to say so and what happens next. The crawler that collects all of this identifies itself and can be blocked in one line.