A new class of AI browsers does not just read the web, it operates it: clicking buttons, filling forms, finishing a task someone asked for. We check whether one of them can reach your page, find the controls, and get through. No score — a verdict, the counts behind it, and one thing to fix.
A page can pass every crawlability and content check on this site and still be impossible to complete a task on.
Reachable URL, HTML that contains the content, structure that says what the page is. It takes what it finds and leaves. Nothing on the page has to respond.
Labelled fields it can identify, controls it can operate without a mouse, states it can tell apart, and a confirmation it can recognise as success. One unlabelled input ends the task.
We report the specific step, not a score. “Book a demo” on a page that reads perfectly:
Most of these are also accessibility problems, so fixing them pays twice.
unlabelled control is enough to end an otherwise complete task
problems fixed at once — agent access and screen-reader access are the same work
trace of the failed attempt in your analytics
Run the check on any URL. You get the step-by-step attempt, the control that ended it, and what to change.
Driving a real browser to finish something a person asked for — book the demo, add to basket, get through the signup. It clicks, types and waits like a visitor would, which means it can be stopped by things a reader would never notice.
No, deliberately. Agentic browsing returns a verdict — Ready, Needs work, Blocked or No action found — and a count, never a number, and it is not one of the dimensions your AI Visibility Score is built from. Not every page should be operable, so scoring every page against it would be wrong.
No. An article, a policy page, a piece of documentation has nothing to operate, and that is the correct answer for it. It is only worth investigating if you expected an action there — which usually means the control appears after an interaction the agent never got to make.
The agents exist and are already in our crawler list — ChatGPT-User, Perplexity-User, Claude-User and GoogleAgent-Mariner among them. Volume is small today. What makes it worth doing early is that the work is the same work as accessibility, so it is not a bet on a timeline.
Three things, in order. Getting in at all — a login wall, a bot challenge, a hard block. Then whether the controls have names it can identify. Then whatever sits in the way: consent overlays, modals, interstitials, and whether each can be dismissed.
No. We walk up to the point of submission and report what an agent could identify and operate. You do not get test leads.
On the crawlability side — that is where getting in is decided. A challenge that stops a crawler almost always stops an agent too, so fix it once and both improve.
They overlap heavily, which is the good news. An accessibility audit measures conformance against a standard. This runs one task end to end and tells you the step it stopped on.
No. A well-built page is already the interface. Real form controls, proper labels and state changes the page announces get you most of the way.
Every paid plan.