# Lighthouse Agentic Browsing: what it checks

Lighthouse, the audit tool built into Chrome DevTools and PageSpeed Insights, has a category for
AI agents: **Agentic Browsing**. It arrived in Lighthouse 13.3.0 in May 2026 and asks a narrower
question than an AI search audit: can an agent that drives a browser understand and operate this
page?

Google labels it experimental: "This category is still under development and subject to change."

## How it scores

Unlike Performance or SEO, the category has no 0–100 score. It shows a fraction, the number of
audits that pass out of those that apply, so a page can show 4/5 rather than 80.

Testing it in DevTools needs Chrome 150 or later. The WebMCP audits need the WebMCP origin trial.
Lighthouse 13.4.1 enabled the category in the PageSpeed Insights API.

## The seven audits

These are the audits in the category in the current Lighthouse configuration.

### Accessibility tree is well-formed

`agent-accessibility-tree`. Agents read the accessibility tree, the same structure screen readers
use. Chrome's documentation calls it their "primary data model". This audit runs a subset of the
accessibility audits in three groups: names and labels, tree integrity, and visibility.

To pass: give every button, link and form field an accessible name, use real `<button>`, `<a>`
and `<label>` elements instead of clickable `<div>`s, keep ARIA roles valid, and do not hide
interactive content from the tree.

### WebMCP tools registered

`webmcp-registered-tools`. Lists the WebMCP tools the page registered when Lighthouse analysed it,
both declarative (forms) and imperative (JavaScript), with their names, descriptions and input
schemas. Lighthouse warns when a page registers more tools than it recommends, because each tool
"consumes model context tokens, adds latency, and increases the risk of tool confusion."

### WebMCP form coverage

`webmcp-form-coverage`. Lists forms without WebMCP annotations and suggests adding them. It passes
when every form is covered.

### WebMCP schemas are valid

`webmcp-schema-validity`. Reports the schema errors the browser found: a form without `toolname`
or `tooldescription`, a field without a `name`, a field without a description. WebMCP and the
attributes are covered in [WebMCP: let AI agents use your forms](/guides/webmcp).

### Cumulative Layout Shift

`cumulative-layout-shift`, the same metric as in Performance. An agent locates an element and
then clicks where it was; if the layout moves in between, it clicks the wrong thing. Reserve
space for images, embeds and banners.

### llms.txt follows recommendations

`llms-txt`. Fetches `/llms.txt`. A missing file is not applicable, because the file is optional.
A server error fails the audit, and so does a file without an H1 heading, a file that is
suspiciously short, or one without links. See [llms.txt: what it is](/guides/llms-txt).

### ARD manifest is valid

`ard-schema`, added in Lighthouse 13.5.0. Agentic Resource Discovery is a proposed manifest at
`/.well-known/ard.json`, or linked with `rel="ard"`, that lists the resources an agent can use: MCP
servers, agents, API descriptions. Lighthouse 13.6.0 follows ARD 0.91, which replaced the older
`/.well-known/ai-catalog.json`. A site with nothing for agents to call does not need one.

## What it does not check

The category is about operating a page in a browser. It does not look at whether AI crawlers may
fetch the page, whether the content is in the HTML without JavaScript, or whether a firewall
blocks AI bots, which decide whether the page reaches AI search at all. Those are covered in
[Is your site agent-ready?](/guides/agent-ready-website-checklist).

## Run it

In Chrome 150 or later, open DevTools, go to the **Lighthouse** panel, tick **Agentic Browsing**
and analyse the page. From the command line:

```sh
npx lighthouse https://example.com --only-categories=agentic-browsing --view
```

A [CrawlWise scan](/) runs axe accessibility checks, measures layout shift, reads `llms.txt`,
records the WebMCP tools a page registers and validates an ARD manifest, alongside the crawler
access and content checks Lighthouse leaves out.

## Sources

- Chrome: [Lighthouse Agentic Browsing](https://developer.chrome.com/docs/lighthouse/agentic-browsing/)
- Chrome: [Lighthouse llms.txt audit](https://developer.chrome.com/docs/lighthouse/agentic-browsing/llms-txt)
- Lighthouse: [Release notes](https://github.com/GoogleChrome/lighthouse/releases)
- Lighthouse: [Default configuration](https://github.com/GoogleChrome/lighthouse/blob/main/core/config/default-config.js)
- Chrome: [WebMCP](https://developer.chrome.com/docs/ai/webmcp)
- ARD project: [Agentic Resource Discovery](https://agenticresourcediscovery.org/)

---

Last updated 4 November 2026.

HTML version: <https://crawlwise.dev/guides/lighthouse-agentic-browsing>

Every CrawlWise page in Markdown: <https://crawlwise.dev/llms.txt>