AI agents
Agentic browsing
Agentic browsing is when an AI agent uses a website the way a person would: reading pages, clicking buttons and filling in forms.
In plain words
The agent works from a machine-readable view of your site, not from a screen. web.dev names three: screenshots, raw HTML and the accessibility tree.
Google’s Lighthouse tool now checks how ready a page is for this.
Example
Lighthouse 13.3, released in May 2026, added a default "agentic browsing" category, with audits for llms.txt and WebMCP. Its release notes said it would ship in the DevTools of Chrome 150 and in PageSpeed Insights.
Why it matters for AI search
An agent that can’t identify a button or a form field can’t finish the task it was given.
Much of what helps agents also helps people who use screen readers, because both rely on the accessibility tree.
What you can do
- Run Lighthouse and look at the agentic browsing category.
- Use semantic HTML: real buttons, links and labelled form fields.
- Make key buttons and fields visible and large enough to be recognized, as web.dev recommends.
Often confused with
- Crawling
- Crawling collects pages. Agentic browsing uses them to complete a task.
- llms.txt
- llms.txt is one file Lighthouse checks in this category. Agentic browsing is the whole idea.
Sources
Facts and sources last checked on 19 October 2026.
Related terms
Go deeper on GEO-Score
How does your page do in AI search?
GEO-Score checks one page on 22 signals, from AI crawler access to structured data, and shows what to fix first.
Analyze a page