Can AI agents actually use your website?
Agents drive real browsers and read the accessibility tree. A site can be perfectly crawlable and completely unusable by one - and most sites have never been tested.
By Sapun Lamichhane · Arcetis
Probably not, and you would have no way of knowing. AI agents that book, buy and fill forms on a user's behalf drive a real browser: they read the accessibility tree, find a control by its name, click it, and check whether the state changed. Nothing about that resembles crawling, and nothing in a normal audit tests it.
Which is how a site ends up perfectly crawlable and completely unusable - ranking well, rendering beautifully, and stopping an agent at the first button.
What stops an agent?
- Controls that are not controls. `<div class="btn" onclick=...>` has no role, no accessible name and no keyboard behaviour. It does not exist in the accessibility tree, so an agent cannot find it. This is the single most common blocker.
- Controls with no name. An icon-only button with no aria-label can only be described by position, which breaks the moment your layout changes.
- Hover-only menus. An agent has no pointer to hover with, so everything behind that menu is unreachable.
- Unlabelled form fields. An agent filling a form matches your labels to what it was asked to enter; an unlabelled field is a guess, and a wrong guess in a checkout is worse than a refusal.
- Consent walls that block the DOM until dismissed. A banner is fine; a modal covering the page stops the agent before it starts.
- Infinite scroll with no pagination. An agent cannot scroll forever, so a listing has an effective depth of one screen.
What makes a site easy for an agent?
Almost everything on the list is something you would do anyway for accessibility. Two are specific to agents:
- `autocomplete` tokens on form fields. `autocomplete="email"` is the closest thing to a machine-readable schema for a form - it tells an agent exactly what a field wants, with no inference from placeholder text.
- schema.org Action markup. `potentialAction` on your Organization or WebSite tells an agent what your site can *do*, not just what it says. That is the difference between an agent reading your restaurant page and booking a table.
Why does this matter commercially?
Because an agent-mediated visit that fails leaves no trace you would recognise. There is no bounce to investigate, no error in your logs, no support ticket. The agent tries, cannot find the control, and reports back to the user that it could not complete the task - and the user hears that your site did not work.
The honest caveat: this is early. Agent behaviour is not specified anywhere - there is no RFC for how an agent picks a button - so rules encoding observed behaviour deserve lower confidence than rules restating WCAG, and should be weighted that way rather than presented as certainty.
How do I test my own site?
The AI agent readiness checker on this site runs 28 rules grouped by what an agent has to do: perceive the page, navigate it, act on controls, complete a transaction, understand what is on offer, and get past whatever is in the way.