Agentic Browsing in PageSpeed Insights: how to pass its audits
Agentic Browsing is Lighthouse's newest category, scored as audits passed, not 0 to 100. Every audit in it, what fails it and the fix for each one.
LLaunchScaler·Published ·8 min read
Agentic Browsing is the Lighthouse category, shown in PageSpeed Insights, that checks whether AI agents can read and operate a page: it audits the accessibility tree agents navigate by, layout stability, your llms.txt file, WebMCP integration and an agent resource catalog. It reports a fraction of audits passed rather than a 0 to 100 score. Every page is judged on its accessibility tree and its layout stability; the others apply only when your server answers /llms.txt with anything but a 400-range status, or you add WebMCP tools or an ai-catalog.json.
The category is new and still changing, so the audit list below is dated: it is what Lighthouse 13.5.0 ran in September 2026.
What is Agentic Browsing in PageSpeed Insights?
It is Lighthouse's category for "browsable websites for AI agents," in the category's own description, which adds that it validates "the correctness of WebMCP integrations" and "is still under development and subject to change." Lighthouse 13.2.0 added it on May 1, 2026, and 13.3.0 moved it into the default configuration on May 7, 2026.
Chrome's documentation marks the category as experimental: "The Agentic Browsing category and WebMCP support are experimental and based on proposed standards." Testing it in Chrome's own DevTools requires Chrome 150 or later, and the WebMCP audits require registering for the WebMCP origin trial. PageSpeed Insights runs it for you: in September 2026, a PageSpeed Insights API request ran Lighthouse 13.5.0 and returned the Agentic Browsing category alongside the usual ones.
Chrome frames the category as a data-gathering exercise: "the current focus is to gather data and provide actionable signals rather than a definitive ranking." Google has not said it affects search rankings.
How is the Agentic Browsing score calculated?
It is a fraction, not a weighted 0 to 100 score. Chrome's documentation says the report shows "a ratio showing how many agentic readiness checks your site passes," plus pass or fail status on specific audits and informational counts. In Lighthouse's configuration, every audit in the category carries a weight of 1 and the category's display mode is fraction.
Audits that do not apply to a page are marked not applicable and drop out of the count. A page with no WebMCP tools and no is not penalised for either. When the PageSpeed Insights API tested example.com in September 2026, the accessibility tree and layout shift audits passed and the other five were not applicable, so the page passed everything that applied.
Questions, answered
What people ask about this
01
What is Agentic Browsing in PageSpeed Insights?
It is a Lighthouse category that checks how well a page is built for AI agents to read and operate: a well-formed accessibility tree, layout stability, an llms.txt file, WebMCP integration and an agent resource catalog. Lighthouse added it to its default configuration in version 13.3, in May 2026.
Chrome lists three reasons a result can change between runs without you changing anything:
WebMCP tools registered with JavaScript may or may not be registered by the moment Lighthouse takes its snapshot.
Significant changes to DOM size or complexity can change how the accessibility tree is built.
Cumulative Layout Shift from ads, images without dimensions or injected content varies from load to load.
Which audits are in the Agentic Browsing category?
In September 2026 the category held seven audits in three groups: Agent Accessibility, Agent Discoverability and WebMCP, plus Cumulative Layout Shift on its own. The table gives each audit's ID, what fails it and the fix, drawn from Chrome's audit pages and the Lighthouse source.
Audit (ID)
Group
Fails when
Fix
Accessibility tree is well-formed (agent-accessibility-tree)
Agent Accessibility
Any of 33 targeted accessibility rules finds a violation, such as a button, link or form field with no accessible name
Use semantic elements and give every control a programmatic name
Cumulative Layout Shift (cumulative-layout-shift)
None
The page shifts enough to fail Lighthouse's CLS audit
Reserve space for images, embeds and injected content
llms.txt follows recommendations (llms-txt)
Agent Discoverability
/llms.txt returns a 500-range error, or loads without an H1, without a link or under 50 characters
Publish a valid file or return a clean 404
ai-catalog.json schema is valid (ard-schema)
Agent Discoverability
An ai-catalog.json exists but does not conform to the Agentic Resource Discovery specification
Validate the file, or do not publish one
WebMCP schemas are valid (webmcp-schema-validity)
WebMCP
A WebMCP form has a tool name without a description or the reverse, or a required field lacks a name
Pair every toolname with a tooldescription and name every input
WebMCP tools registered (webmcp-registered-tools)
WebMCP
Informational; warns above 40 registered tools
Register only the tools the current page needs
WebMCP form coverage (webmcp-form-coverage)
WebMCP
Informational; lists forms without WebMCP attributes
Add toolname and tooldescription to forms agents should use
The ard-schema audit is the newest; Lighthouse 13.5.0 added it on September 18, 2026 and grouped it with llms.txt under Agent Discoverability. Expect the list to keep growing.
How do you pass the accessibility tree audit?
Make every interactive element nameable by software. Chrome's documentation says agents "review the accessibility tree to identify interactive elements," and "missing labels can block users with visual disabilities and agents from completing a task." The audit runs a targeted set of accessibility rules and fails on any violation.
The rules the audit checks include button-name, link-name, label, input-button-name, select-name, document-title, svg-img-alt, tabindex, autocomplete-valid, duplicate-id-aria and a long list of aria-* rules. Those rules map onto a handful of common patterns:
Pattern that fails
What an agent sees
Fix
<button><svg>...</svg></button> with no text
A button with no name
Add visible text or aria-label="Close"
<a href="/pricing"><img src="arrow.svg"></a> with no alt
A link with no name
Add link text, or alt text on the image
<input type="email" name="email"> with no label or aria-label
A field with no name
Add <label for="email">Email</label> and give the input id="email"
Repeated id values used by ARIA references
References that point at the wrong element
Make every referenced id unique
A page with no <title>
An unnamed document
Add a descriptive <title>
These are the same fixes an accessibility audit asks for, because agents read the page the way assistive technology does. The guide to making a website usable by AI agents covers the wider set of things that break an agent session.
How do you pass the layout stability audit?
Stop elements moving after they appear. Chrome's documentation says agents "often rely on screenshots or coordinate-based interaction," so "unexpected layout shifts can cause an agent to miscalculate the position of a button or input." The audit reuses Lighthouse's Cumulative Layout Shift measurement.
web.dev's threshold for a good CLS is 0.1 or less, and anything above 0.25 is poor. The usual causes and fixes:
Images and videos without dimensions: set width and height attributes, or reserve the box with the CSS aspect-ratio property.
Ads, embeds and iframes that resize: give their container a fixed minimum height before they load.
Content injected above existing content, such as banners and cookie bars: reserve its space in the initial layout, or place it where it does not push content down.
Web fonts that render at a different size from the fallback: match the fallback's metrics so the swap does not reflow text.
Either publish a valid file or make /llms.txt return a clean 404. The audit is marked not applicable when the file is missing, because Chrome says "providing the file is optional at the moment." It fails only when the URL returns a server error, or when the file loads but lacks an H1, a Markdown link or 50 characters.
The Lighthouse source checks three things once the file loads: a line starting with # followed by text, at least one link written as [text](url), and a length of at least 50 characters. A two-line file with a heading and a link to your docs passes. Two failure modes are easy to miss: a catch-all route that answers /llms.txt with a 500, and one that answers it with your HTML home page, which has no Markdown H1 or Markdown link. The guide to llms.txt has a full example.
Do you need WebMCP or an ai-catalog.json?
Not to pass the category today. With no WebMCP tools and no ai-catalog.json, those audits are not applicable and do not count against you. WebMCP is an experimental way to expose your site's actions to agents directly, and it requires the origin trial, so add it only if you want agents to complete tasks such as bookings through defined tools.
If you do adopt it, the declarative form is attributes on a normal HTML form, as Chrome's documentation shows:
<form toolname="newsletter_signup" tooldescription="Subscribes the user to the weekly newsletter">
<input name="email" type="email" toolparamdescription="The user's email address">
<button type="submit">Sign Up</button>
</form>
The schema audit fails a form with a toolname but no tooldescription, or the reverse, and a required field with no name. It warns when an optional field has a name but neither a toolparamdescription nor a <label>. Chrome also advises against registering more than 40 tools at once.
How do you run the Agentic Browsing audit?
Run PageSpeed Insights on the URL, or run Lighthouse yourself. From the command line, npx lighthouse https://yoursite.com --only-categories=agentic-browsing --view runs just this category. In Chrome DevTools, the category needs Chrome 150 or later. The PageSpeed Insights API accepts category=AGENTIC_BROWSING in the request.
Run it on the pages agents are most likely to use, such as pricing, signup and contact, not only the home page. Then re-run after each fix; the accessibility and layout audits change with your templates, so one template fix often clears the same failure across the site.
See the Agentic Browsing factor next to your full readiness scan
Lighthouse checks how a page is built; it does not check whether an AI agent is allowed in. Run the free scan on LaunchScaler: it takes a URL, needs no account and runs 156 checks across 6 of its 7 categories at no cost, and it reports an Agentic Browsing factor out of 3 alongside those checks. The factor rolls up three things an agent needs. Access asks whether ChatGPT-User and Claude-User get a real page rather than a challenge or an empty shell, and whether robots.txt and your firewall let the retrieval bots in. Comprehension covers form labels, link and heading semantics, image alt text, the page language and llms.txt. Operability covers dead calls to action, keyboard operability, cookie banners that block the page and hydration errors. A dimension counts only when every signal in it passes.
How is the Agentic Browsing score calculated?
It is not a 0 to 100 score. Lighthouse shows a fraction: how many of the category's applicable audits the page passes. Audits that do not apply, such as WebMCP checks on a page with no WebMCP tools, are marked not applicable.
03
Which audits are in the Agentic Browsing category?
In September 2026 the category held seven: agent accessibility tree, Cumulative Layout Shift, llms.txt, ai-catalog.json schema validity, and three WebMCP audits covering form coverage, registered tools and schema validity.
04
Does the Agentic Browsing score affect Google rankings?
Google has not said it does. The category is marked as still under development and subject to change, and Chrome's documentation describes its focus as gathering data and giving actionable signals rather than a definitive ranking.
05
Why does my Agentic Browsing result change between runs?
Chrome's documentation lists three causes: WebMCP tools registered by JavaScript may or may not be captured in the snapshot, large changes to the DOM can change the accessibility tree, and layout shifts move elements between runs.
Every major AI crawler user agent and robots.txt token, with its operator, purpose, JavaScript rendering and published IP list, checked September 2026.
Vercel measured ChatGPT's crawler spending 34.82% of fetches on 404s and 14.36% on redirects, against 8.22% and 1.49% for Googlebot. How to cut the waste.
When ChatGPT names competitors and not you, list who wins each prompt, open the sources it cited, and get onto those pages. A step-by-step gap analysis.
