What Is Agentic Browsing, and Why Should a Business Care?

How is this different from an AI chatbot?

A chatbot answers from what it already knows, or from a search result it was handed. An agent goes and does the task. It opens pages, works through them, and takes the next step.

That difference is the whole story. A chatbot never touches your site, so your site's structure is irrelevant to it. An agent is a visitor — one that navigates entirely by structure, with the visual layer stripped away.

What actually breaks a site for an agent?

In August 2026 we ran Google's Lighthouse agentic checks across 91 live business websites. 52 passed everything. 39 had at least one defect that would obstruct an agent. The failures clustered tightly.

Links with no readable text — the most common problem by far

Twenty of the 91 sites had links an agent cannot identify. Usually an icon with no label, or a bare arrow. A person sees a chevron and understands "next". An agent sees a link that goes somewhere unnamed, and has no basis for following it.

Unlabelled form fields

One site had four contact-form fields with no labels attached. A customer infers what goes where from the layout. An agent has nothing to infer from — it cannot tell which box is the phone number. If you want an agent to be able to submit an enquiry on someone's behalf, that form has to be labelled.

Elements hidden from agents but still clickable

A recurring pattern where something is marked as hidden from assistive technology while remaining interactive on screen. Humans click it; agents skip it entirely. If that element is your "Request a quote" button, the agent simply never finds the path.

Pages that move while they load

Twelve sites failed on layout stability. Content shifts as images and ads settle. Sighted users adjust. An agent working from coordinates clicks where the button was a moment ago.

What does this have to do with accessibility?

Almost everything, and this is the most reassuring part.

The agentic checks lean heavily on the accessibility tree — the same structural layer a screen reader uses. Unlabelled links, unlabelled fields and misapplied ARIA attributes have been accessibility failures for years. An AI agent is, functionally, a very literal non-visual user.

So the work is not speculative. Fixing these makes your site usable for people using assistive technology, which is worth doing regardless of whether a single agent ever visits.

Should you act on this now?

A straight answer: agentic browsing is a small share of traffic today. We are not going to tell you it is urgent, and anyone quoting you a precise figure for how much traffic it represents is guessing.

The case for looking now is proportion, not panic. Most of these defects are twenty-minute fixes. They improve accessibility immediately, they overlap with technical SEO, and they remove a category of failure you currently cannot see. The cost of checking is an afternoon. The cost of finding out later is that an agent quietly failed to use your contact form for a year.

How do you check?

Google's Lighthouse includes agentic browsing audits covering the accessibility tree, layout stability and llms.txt. Our free agent-readiness check runs those same rules against any public URL and reports the results in plain English.

One honest caveat about scoring: Lighthouse defines four checks, but one of them — WebMCP — requires browser-level inspection that cannot be run remotely. Any remote tool, ours included, is reporting on three. We say so on the results rather than quietly scoring out of four.

Being usable by an agent is a separate question from whether AI mentions you at all. For that, see how to get cited by AI.