Lighthouse Agentic Browsing audits: what each check tests
What each Lighthouse Agentic Browsing audit tests: agent accessibility, CLS, llms.txt, ARD and WebMCP, how to run the category alone, and how to fix failures.
By Crawlito · 9 min read
In short
- The Agentic Browsing category is experimental, needs Chrome 150 or later, and shows a pass ratio instead of a weighted 0 to 100 score; not applicable and informative audits are left out of it.
- Its agent accessibility audit fails if any of 33 axe rules fails, including button-name, link-name, label, aria-hidden-focus and aria-required-children.
- In the Lighthouse source, the llms.txt audit fails on a 5xx or 429 response and also on a file with no H1, no Markdown links or fewer than 50 characters; a 404 is not applicable.
- The WebMCP audits return not applicable when the browser does not expose WebMCP; Chrome says they need the WebMCP origin trial, and its toolkit post tests them with a Chrome flag.
- Passing the category is a browser agent check: Chrome says ordinary SEO principles still apply when agents only search, so a pass says nothing about Google Search or AI answers.
On this page
- What the Agentic Browsing category is
- How to run only the Agentic Browsing category
- The agent accessibility audit: names, labels and tree integrity
- Cumulative Layout Shift in the agent category
- The llms.txt and ARD discovery checks
- The WebMCP audits
- What a pass does and does not tell you
- Frequently asked questions
- Next step
A Lighthouse Agentic Browsing audit checks how easily an AI agent running in a browser can read and operate one page. The category groups five kinds of check: an agent-focused subset of the accessibility audits, Cumulative Layout Shift, an llms.txt check, an Agentic Resource Discovery (ARD) manifest check and three WebMCP audits. It is experimental, needs Chrome 150 or later, and reports a pass ratio rather than a 0 to 100 score. Audits that are not applicable or informative are left out of that ratio, so on a site with no llms.txt, ARD manifest or WebMCP, the accessibility audit and CLS usually decide the result.
This post goes through each audit as the Lighthouse source code defines it, because the Chrome documentation describes the checks in broad terms. Facts were checked against Lighthouse 13.6.0 and the main branch on 15 November 2026. For what llms.txt is and how to write one, see the llms.txt guide; for WebMCP itself, see WebMCP explained.
What the Agentic Browsing category is
Lighthouse added the category in version 13.2.0 on 1 May 2026 and put it in the default config in 13.3.0 on 7 May. Chrome's toolkit post of 22 June 2026 says it is available from Chrome 150. The category description in the source says it is "still under development and subject to change", and Chrome's docs call it experimental and based on proposed standards.
Scoring works differently from the other categories. Chrome says there is no weighted 0 to 100 score: the header shows a fraction of checks passed, and each audit shows pass or fail, sometimes with warnings. In the default config, the category is set to categoryScoreDisplayMode: 'fraction' and lists seven audits:
| Audit ID | Group | What it tests | Typical result on an ordinary site |
|---|---|---|---|
agent-accessibility-tree | Agent Accessibility | Names, labels, ARIA validity, hidden focusable content | Pass or fail |
cumulative-layout-shift | none | Visual stability during load | Pass or fail |
llms-txt | Agent Discoverability | /llms.txt fetch and basic format | N/A if there is no file |
ard-schema | Agent Discoverability | ARD manifest at /.well-known/ard.json or linked | N/A if there is no manifest |
webmcp-registered-tools | WebMCP | Lists tools the page registered | N/A without WebMCP support |
webmcp-form-coverage | WebMCP | Forms without WebMCP annotations | N/A without WebMCP support |
webmcp-schema-validity | WebMCP | Errors in declared tool schemas | N/A without WebMCP support |
The two WebMCP listing audits are informative, so they report findings without passing or failing. In the report renderer, not applicable and informative audits are skipped when the fraction is counted, and an audit counts as passed at a score of 0.9 or more. That threshold matters for the audits that give partial scores: an ARD manifest with only warnings (0.9) counts as passed, while WebMCP schema warnings (0.5) do not.
Chrome's own page summarises the accessibility part as three themes: names and labels for interactive elements, tree integrity, and visibility in the accessibility tree.
How to run only the Agentic Browsing category
In Chrome 150 or later, open DevTools, go to the Lighthouse panel, select only the Agentic Browsing category and run the report. From the command line, the Lighthouse CLI accepts the category ID:
npx lighthouse https://example.com/ \
--only-categories=agentic-browsing \
--output=html --output-path=./agentic.htmlThe CLI needs Node 22 or later, according to the Lighthouse readme (the 13.x package sets 22.19 as the minimum). To run a single audit, use --only-audits=agent-accessibility-tree. The category supports navigation and snapshot modes, but the WebMCP audits only run in navigation mode, and the Chrome toolkit post describes testing WebMCP through Chrome Canary with chrome://flags/#enable-webmcp-testing enabled. Chrome's scoring page says the WebMCP audits require the WebMCP origin trial, which opened in Chrome 149.
Expect some variation between runs. Chrome lists three causes: the timing of dynamic tool registration, DOM changes that alter the accessibility tree, and layout shifts from ads, unsized images or injected content.
The agent accessibility audit: names, labels and tree integrity
The agent-accessibility-tree audit does not run its own checks. It takes the page's axe-core results from the normal accessibility gatherer and fails if any rule in a fixed list fails. The audit source lists 33 rule IDs. The source keeps them in one flat list; the grouping below is ours, to make it easier to read:
| Theme | axe rules in the list |
|---|---|
| Names and labels | button-name, input-button-name, input-image-alt, label, link-name, select-name, svg-img-alt, document-title, aria-command-name, aria-dialog-name, aria-input-field-name, aria-toggle-field-name, aria-tooltip-name, aria-treeitem-name |
| Tree integrity | aria-roles, aria-allowed-role, aria-required-children, aria-required-parent, aria-required-attr, aria-allowed-attr, aria-conditional-attr, aria-prohibited-attr, aria-valid-attr, aria-valid-attr-value, presentation-role-conflict, definition-list, duplicate-id-aria, table-duplicate-name |
| Visibility and focus | aria-hidden-body, aria-hidden-focus, aria-text, tabindex |
| Forms | autocomplete-valid |
The result is binary. One failing rule fails the audit, and the report shows a Failed Audits section listing each failing rule's help text and the first failing element. Colour contrast, heading order and landmark rules are not in the list, so they appear only in the regular Accessibility category.
The reason is practical. Chrome's page on accessibility for agents says agents review the accessibility tree to identify interactive elements, and that missing labels can block both agents and users with visual disabilities from completing a task.
How to fix the common failures
Most failures are an icon button, an unlabelled input or an ARIA role used without its required children. The WAI-ARIA Authoring Practices Guide covers the correct patterns. Before and after for the three most frequent cases:
<!-- Before: no accessible names, fails button-name and label -->
<button class="icon-search"><svg>...</svg></button>
<span>Email</span>
<input type="email" name="email">
<!-- After: a name the accessibility tree can expose -->
<button class="icon-search" aria-label="Search"><svg aria-hidden="true">...</svg></button>
<label for="email">Email</label>
<input id="email" type="email" name="email" autocomplete="email"><!-- Before: fails aria-hidden-focus, a hidden region still holds a focusable link -->
<div aria-hidden="true"><a href="/offer">Offer</a></div>
<!-- After: hide it from everyone, or not at all -->
<div hidden><a href="/offer">Offer</a></div><!-- Before: fails aria-required-children, a menu with no menuitem children -->
<ul role="menu"><li><a href="/pricing">Pricing</a></li></ul>
<!-- After: plain navigation needs no ARIA role -->
<nav aria-label="Main"><ul><li><a href="/pricing">Pricing</a></li></ul></nav>Note that axe accepts a non-empty placeholder as a label for the label rule, so an input with only a placeholder passes this audit; a real <label> is still the better fix, because WebMCP and assistive technology both read it. In the last case, a site menu marked up as role="menu" promises application-menu behaviour it does not have. Native elements are usually the fix. Page-level structure such as headings and landmarks is a separate topic, covered in headings, landmarks and link text.
Cumulative Layout Shift in the agent category
The category reuses the standard cumulative-layout-shift audit, the same one the Performance category uses. Lighthouse scores it on a curve with control points of 0.1 (a score of 0.9) and 0.25 (a score of 0.5). Since the fraction counts an audit as passed at 0.9 or more, a lab CLS of about 0.1 or lower is what passes here. Chrome's overview says visual stability is critical for agents that rely on element positioning.
The usual fixes apply: give images and embeds width and height or an aspect-ratio, reserve space for ad slots and cookie banners, and avoid inserting content above what is already on screen. In navigation mode the CLS value comes from a lab page load, not from real users; the Core Web Vitals guide explains field data and what page speed does and does not change for crawlers.
The llms.txt and ARD discovery checks
The llms-txt audit fetches /llms.txt from the root of the page's origin. Chrome's llms.txt audit page says it flags a server error and marks a missing file as not applicable. The audit source checks more than that:
| Response or content | Result |
|---|---|
| 5xx, or 429 from Lighthouse 13.6.0 | Fail, "Failed with HTTP status ..." |
| Other 4xx, including 404 | Not applicable |
| Network error or no status | Fail, "Fetch of llms.txt failed" |
200 but no line starting with # (leading spaces allowed) | Fail, missing H1 |
200 but no [text](url) link | Fail, no links |
| 200 but under 50 characters | Fail, suspiciously short |
| 200 with an H1, a link and 50 characters or more | Pass |
The common surprise is a 200 response that is not the file at all. Some single-page apps and CDNs answer every unknown path with the HTML shell or a challenge page. Lighthouse then reads HTML, finds no Markdown H1, and fails a file you never published. Return a real 404 for /llms.txt if you have none:
curl -s -o /dev/null -w '%{http_code} %{content_type}\n' https://example.com/llms.txtThe ard-schema audit, added in 13.5.0 and moved to the ARD v0.91 format in 13.6.0, looks for an Agentic Resource Discovery manifest, which Chrome describes as listing resources such as MCP servers, A2A agents and agent skills. It checks an Agentmap: line in robots.txt, a <link rel="ard"> tag, a Link header with rel="ard" and /.well-known/ard.json, then validates the JSON against a port of the ARD v0.91 conformance test. The older ai-catalog locations are still read, with a warning. With no manifest anywhere it is not applicable and does not affect the result.
The same trap as llms.txt applies. In the audit source, the not applicable result needs no advertised manifest and a 4xx from /.well-known/ard.json. A 5xx there fails the audit, and an HTML page served with a 200 is validated as a manifest and fails as malformed JSON. A manifest you advertise that cannot be loaded also fails. Publish one only if you actually offer agent resources.
The WebMCP audits
All three WebMCP audits return not applicable when the browser does not support WebMCP, so on stable Chrome without the origin trial or testing flag you will normally see N/A. When support is present:
webmcp-registered-toolslists imperative tools (registered from JavaScript, with their source location) and declarative tools (forms), with each input schema. It warns above 40 tools, saying more consume model context and raise the risk of tool confusion.webmcp-form-coveragelists forms that have neither atoolnamenor atooldescriptionattribute. It is informative and not applicable on a page with no forms; since 13.5.0 it shows as a pass when every form is annotated.webmcp-schema-validityreads the browser's schema issues. A missingtoolname, a missingtooldescriptionor a required field with nonameis an error and scores 0. A field with no title or description, or an optional field with no name, is a warning and scores 0.5, which the fraction counts as not passed. With no tools and no issues it is not applicable.
Chrome's declarative API page shows the attributes. A form that clears the schema audit looks like this:
<form toolname="supportRequestTool" tooldescription="Submit a request for support." action="/support">
<label for="team">Team</label>
<select id="team" name="team" required toolparamdescription="Which team handles the request.">
<option value="billing">Billing</option>
<option value="technical">Technical</option>
</select>
<button type="submit">Send</button>
</form>Chrome says a field without toolparamdescription falls back to its <label> text, and to aria-description when there is no label, which is one more reason the accessibility fixes come first. Chrome also notes that removing either toolname or tooldescription unregisters the tool.
What a pass does and does not tell you
The category tests one page in one browser for agents that act on it. It does not test whether crawlers can reach the page, whether the content is in the raw HTML, or anything a search engine indexes. Chrome's toolkit post says that when agents only search for websites, the existing SEO principles still apply, and Google's AI optimization guide says Google Search does not use llms.txt or other AI text files. No search operator documents using these audits as a ranking or citation signal, so treat a pass as a usability check for browser agents, not as a route into AI answers.
It also overlaps only partly with other readiness checks. Cloudflare's scanner looks at site-wide files and headers, as compared in the Cloudflare agent readiness score post, and Lighthouse's SEO category has its own robots.txt audit, explained in robots.txt is not valid in Lighthouse.
Frequently asked questions
Why are my WebMCP audits all not applicable?
Lighthouse marks them N/A when the browser does not support WebMCP. Chrome says the WebMCP audits need the origin trial, and its toolkit post tests them in Chrome Canary with the enable-webmcp-testing flag.
Does a missing llms.txt lower my Lighthouse score?
No. A 404, or any other 4xx response except 429, makes the audit not applicable, and not applicable audits are left out of the pass ratio. It fails on a server error, a 429, or a file that lacks an H1, a Markdown link or 50 characters.
Is the Agentic Browsing score part of the Performance or SEO score?
No. It is a separate category with a pass ratio and no weighted score, although it reuses the same CLS audit as Performance. Chrome describes it as experimental, and the Lighthouse source says it is subject to change.
Will passing Agentic Browsing help my site appear in AI search answers?
Nothing in Chrome's documentation says so. The audits test how a browser agent operates a page, while AI search depends on whether crawlers can fetch and read your HTML, and no operator documents using these audits.
Next step
Fix the accessibility rules Lighthouse lists, check that /llms.txt returns a real 404 or a real file, and leave the WebMCP audits until you join the origin trial. To see whether AI crawlers can fetch and read the same page, run the free check on it; it needs no account.
Filed under Rendering and content. Topics: ai-agents, accessibility, llms-txt, audits