Agentic Browsing Audit: Fix Order for All 6 Checks

A quote request form works for every visitor and still fails for an AI agent. The agent finds an unlabeled field, a "button" built from a <div>, and a banner that pushes the form down mid-click. The task ends unfinished.

Lighthouse now reports these problems in a new Agentic Browsing category. This post covers only those six checks, the order to fix them, and what each one changes for search engine optimization (SEO).

The category checks the accessibility tree, Web Model Context Protocol (WebMCP) form coverage, WebMCP schema validity, Cumulative Layout Shift (CLS), llms.txt, and ai-catalog.json. Fix the accessibility tree and layout shift first, because both affect real visitors today. Add the emerging-standard items afterward: a missing file is marked not applicable, and a malformed one fails.

What the six Agentic Browsing audits measure

Lighthouse 13.3 added the Agentic Browsing category in May 2026, and Lighthouse 13.5 added the catalog audit on 18 September 2026. Chrome's documentation labels the category experimental and describes the result as a pass ratio. Some third-party tools convert it into a number, so compare scores only within one tool.

Run it on Chrome 150 or later (Chrome Canary works), because Lighthouse reads WebMCP through the Chrome DevTools Protocol. The audits are deterministic, so the same page returns the same result unless tool registration timing changes.

Weights matter for reading the ratio. ContextBolt's breakdown of the report JSON shows the accessibility tree, CLS, and llms.txt audits at weight 1 and the WebMCP audits at weight 0. The release notes do not document the weight of the catalog audit, so read the weight values in your own JSON report for your Lighthouse version.

Fix the accessibility tree and layout shift first

These two audits describe problems that harm human visitors today, so they come before any new file or protocol. Start with accessibility because the report names each failing element. Layout shift comes second because its fixes touch templates and ad slots.

Lighthouse runs a filtered set of accessibility audits for this check, and agents that work from the accessibility tree depend on the same roles and names that screen readers use. When one template fails, a website accessibility audit finds the same pattern on every page that shares it.

In the quote form scenario, the usual miss is a submit control built from a <div> with a click handler and no role. Replace it with a real <button>, tie a visible <label> to each input, and add aria-label only where no visible text exists.

Layout shift matters for agents because an element can move between the moment it is located and the moment it is clicked. In the same scenario, a cookie banner that loads late pushes the form down by its own height.

Reserve space with width and height attributes or a CSS aspect-ratio, and load banners into a fixed slot. Core Web Vitals treats a CLS of 0.1 or lower as good.

Set up llms.txt and ai-catalog.json files

Both file audits report not applicable when the file does not exist and fail when the file exists but is malformed. Skipping a file costs nothing in the audit, and a broken file costs a failure. Publish a file only when it is correct and useful.

The llms.txt file is a markdown file at the domain root: a heading with the site name, a one-line summary, and sections of links with short descriptions to the most important pages. Chrome lists the exact checks in its llms.txt audit reference.

A GitHub issue on the Lighthouse repository (number 17082) reports a "fetch failed" message for a file that served correctly to two successful requests. Confirm the file returns HTTP 200 as plain text, then check CDN and bot-protection logs for the Lighthouse request before editing the content.

The Agentic Resource Discovery (ARD) audit looks for a catalog pointer first (such as a link tag with the ai-catalog relation) and falls back to /.well-known/ai-catalog.json. Schema errors fail the audit, and so does a pointer to a catalog that does not load.

Publish a catalog only when real tools or services exist for an agent to use. A catalog that lists nothing real passes validation and still gives agents wrong information.

Annotate forms and validate WebMCP schemas

WebMCP lets a page register actions as named tools with defined inputs, so an agent can call the action directly. Lighthouse checks whether forms carry the annotations and whether those annotations match the expected schema. It also lists every tool registered on the page, through declarative form attributes or through navigator.modelContext.registerTool in JavaScript.

Start with the single form that matters most. In the quote form scenario, annotate the quote request declaratively, then confirm that Lighthouse lists the tool and that every input has a schema that parses. An invalid schema is worse than a missing one, because the tool registers and the agent's call then fails.

Tools registered in JavaScript after Lighthouse takes its snapshot do not appear in the report. Register them early in page load, and rerun the audit three times before changing code. If every WebMCP audit reports not applicable, switch to Chrome 150 or later first, then check whether the page has any forms or tools.

Skip WebMCP on pages with no forms or actions. Informational pages have nothing to expose, and the effort belongs on conversion pages first.

How the audits affect SEO and performance

None of the six audits is documented as a Google Search ranking factor. The category sits outside Lighthouse's SEO category, and the 13.5 release notes do not link the catalog audit to Search. Google's AI Optimization Guide (May 2026) says llms.txt is not needed for Google Search, and no major AI provider has publicly confirmed that the file changes how it sources or cites content.

Two audits overlap with signals search engines already read. CLS is a Core Web Vitals metric, and the accessibility checks match the semantic HTML and labeling that help crawlers parse a page. A Core Web Vitals optimization pass on shared templates improves both the Core Web Vitals report and this one.

The performance cost of the other four items is small. A text file, a JSON file, and a few form attributes add bytes without adding render work. The cost sits in maintenance: someone must update the files when pages and tools change.

What happens when an audit fails

The consequence depends on the audit. Failures in the first two affect visitors now, and failures in the file and protocol audits affect only agents that read those files.

• Accessibility tree fails: agents and screen reader users cannot identify controls, so form submissions and checkout steps break.
• CLS fails: targets move during interaction, which causes misclicks for visitors and agents and weakens the Core Web Vitals result.
• Malformed llms.txt or catalog: the audit fails and agents receive unreadable or wrong information. Remove the file until it validates.
• Invalid WebMCP schema: tools register but arguments fail to parse, so an agent cannot call them.

Skipping the emerging standards entirely carries no penalty today. The risk is falling behind if agent-driven booking and buying grows, and that risk is a forecast with no measured ranking effect.

Run the six checks in this order

1. Run Lighthouse on Chrome 150 or later against the page that holds the main form, with the Agentic Browsing category selected.
2. Fix every accessibility tree failure: label each field and replace fake buttons with real ones.
3. Fix layout shift on the same page by reserving space for banners, images, and embeds.
4. Validate llms.txt and the catalog if either file is published, or leave both out.
5. Annotate the main form with WebMCP, check the schema, and rerun the audit three times.

Add Lighthouse to continuous integration after the first clean pass, so template changes do not bring failures back. Recheck the audit list after each Lighthouse release, because the standards behind it are still proposals.

If the failures trace back to shared templates across the site, a technical SEO audit maps them in one pass.

FAQs

People Also Ask