Skip to content
SeenLens

Lighthouse agentic browsing: how to fix failing audits

Lighthouse has an experimental Agentic Browsing category that scores how well browser agents can read and use a page. The score is a fraction of passed audits, not 0 to 100. SeenLens does not run these audits yet and lists the check under not tested. You can run them yourself in one command and fix each failure with the guides below.

By SeenLens · Last updated

Run the free 43-check scan on your page

What is the Lighthouse Agentic Browsing category?

Chrome's documentation describes Agentic Browsing as an experimental category based on proposed standards. It does not use a weighted 0 to 100 score. It shows a fraction: the number of agent readiness audits passed out of the audits that apply. Testing it needs Chrome 150 or later, and the WebMCP audits need an origin trial registration.

In Lighthouse 13.6.0, the latest npm release on 2026-11-06, the category holds seven audits with equal weight.

Audits in the Agentic Browsing category (Lighthouse 13.6.0)
AuditWhat it testsSeenLens guide
agent-accessibility-treeInteractive elements have names, the roles and parents in the accessibility tree are valid, and interactive content is not hidden from the treeSee the accessibility tree section below
webmcp-registered-toolsThe page registers WebMCP toolsAdd WebMCP tools
webmcp-form-coverageForms carry WebMCP annotationsAdd WebMCP tools
webmcp-schema-validityAnnotated forms have a valid tool schemaAdd WebMCP tools
cumulative-layout-shiftLayout does not move after loadCore Web Vitals
llms-txtA server error on /llms.txt fails; a 404 is not applicable; an existing file needs an H1 and a linkAdd llms.txt
ard-schemaThe ARD manifest at /.well-known/ard.json or linked with rel="ard" is validai-catalog.json errors

Lighthouse leaves an audit that is not applicable out of the fraction. A site with no llms.txt and no ARD manifest is therefore scored on the other audits. Chrome's documentation says the category is experimental, so the audit list can change between releases.

SeenLens treats this check as external. The scan lists it under 'not tested' with the reason that the data is not available from the PageSpeed Insights API. SeenLens does test related items itself: llms.txt, ai-catalog.json, WebMCP markup and server rendering.

How do I run the audits?

Use the Lighthouse command line tool. The category id is agentic-browsing.

Terminal
npx lighthouse@latest https://example.com/ --only-categories=agentic-browsing --output=html --output-path=./agentic.html

You can also open Lighthouse in Chrome DevTools and select the category if your Chrome version lists it. Open the report and expand each failed audit. The report shows the elements or messages behind each failure.

How do I fix the accessibility tree audit?

This audit reads the page the way an agent reads it: through the accessibility tree. Chrome lists three tests: names and labels, tree integrity and visibility.

  • Give every button, link and form field a name. Use visible text, a label element or aria-label. An icon button with no text fails.
  • Use native elements (button, a, input) instead of clickable div elements, so roles and parents are valid.
  • Do not hide interactive content from assistive technology with aria-hidden="true" or display: none while it still works on screen.
HTML
<!-- Before -->
<div class="icon" onclick="openCart()"></div>

<!-- After -->
<button type="button" aria-label="Open cart" onclick="openCart()">
  <svg aria-hidden="true">...</svg>
</button>

How do I fix the other audits?

  • WebMCP audits: add tools or annotations only where an agent can safely act. See Add WebMCP tools. If the audit says no tools, and your site has no actions to offer, accept the result.
  • Layout shift: set width and height on images and embeds and reserve space for late content. See Core Web Vitals failing.
  • llms.txt: return a valid file or a real 404. See Add llms.txt.
  • ARD manifest: serve it at /.well-known/ard.json or link it with rel="ard", or return a 404 if you have nothing to list. See ai-catalog.json errors.

How do I verify the fix?

Run the Lighthouse command again and compare the fraction. A higher fraction after the fix confirms it. Keep the Lighthouse version in your notes, because the category is experimental and new releases change audits.

Should I chase a full fraction?

No. The category measures readiness for agents, and Google says its Search features need no special optimization. A fraction below full is expected on a content site with no forms. Fix the audits that match real features of your site, and leave WebMCP and ARD alone if you have no service to expose. SeenLens gives this check weight 3 in the agent readiness dimension, and it is left out of the score while it shows as 'not tested'.

Scan your page with SeenLens

Sources