Start
On this page

Lighthouse agentic browsing: what the new audit checks

A practical guide to Lighthouse's experimental Agentic Browsing category, WebMCP checks, llms.txt, accessibility tree issues, layout stability, and where Layzr goes further.

Updated May 21, 2026

Lighthouse has a new experimental category called Agentic Browsing. It is built around a simple question: can an AI agent understand and interact with your website without getting tripped up by missing labels, shifting layouts, unclear forms, or missing machine-readable context?

This is early. Chrome's own Lighthouse agentic browsing scoring docs say the category and WebMCP support are experimental and based on proposed standards. That makes the feature worth watching, but it also means you should not treat the score like a final ranking system.

The better way to read it: Lighthouse is starting to measure parts of the web that matter to agents. Layzr already looks at the page from the other side of the problem: can AI systems understand the brand, can they mention it in answers, and can an agent-style workflow move through the page? Use both if you care about answer engine optimization, AI visibility, and agent-ready websites.

What is Lighthouse agentic browsing?

Lighthouse agentic browsing is an experimental Lighthouse category for checking whether a website is built in a way machines can interact with. It looks at deterministic signals, not whether a real customer would finish a task.

DebugBear's agentic browsing article notes that the category started appearing with Lighthouse 13.3 and includes checks around the accessibility tree, content shifts, WebMCP, and llms.txt. That matches the direction in Chrome's docs: agents need cleaner structure, stable pages, and clearer machine-facing instructions.

How the agentic browsing score works

The first thing to know: the agentic browsing score is not a normal Lighthouse score. Chrome says the category does not use a weighted 0 to 100 average. It reports a fractional pass ratio, pass or fail audit status, and informational counts.

  • A fractional score shows how many agentic readiness checks passed.
  • Individual audits can show pass, fail, warning, or informational results.
  • The category is meant to provide signals while standards are still forming.
  • Results can move if your WebMCP tools register at different times, the accessibility tree changes, or layout shifts affect interaction targets.

That matters if you already use PageSpeed Insights or Lighthouse for performance work. A 100 Performance score and a clean agentic browsing report are measuring different things. One asks whether the page loads well. The other asks whether an agent can read and use enough of the page structure.

What Lighthouse checks for agents

The category mixes newer agent-specific checks with older website quality signals that now matter more because agents read pages differently from humans.

WebMCP checks

WebMCP is a proposed way for a website to expose actions and form intent to AI agents. Chrome says Lighthouse monitors WebMCP tool registration through the Chrome DevTools Protocol and checks both declarative tools in HTML and imperative tools registered with JavaScript.

One current audit looks for forms that do not include both `toolname` and `tooldescription`. Chrome marks that specific audit as informational for now, so it can tell you what is missing without failing the whole page.

Accessibility tree checks

AI agents often rely on the accessibility tree because it gives them roles, names, states, and relationships without making them parse every visual detail. If your buttons are unlabeled, your custom controls hide their state, or a clickable `div` has no role, the agent has to guess.

  • Use real buttons and links when the element is interactive.
  • Give form fields attached labels.
  • Make names and states programmatic rather than just visual.
  • Avoid hiding interactive elements from the accessibility tree.

Layout stability checks

Layout shift is already part of Core Web Vitals, but it has a different cost for agents. If an agent identifies a button and the page moves before the click, the action can hit the wrong target.

This is why the new category pays attention to CLS. The fix is the same as normal page speed work: reserve image and embed space, avoid injecting banners above content, load fonts carefully, and keep important controls stable.

llms.txt checks

The llms.txt Lighthouse audit checks for a machine-readable Markdown summary at the site root. Chrome says a missing file returns N/A for now because llms.txt is optional, while server errors can be flagged.

If you add one, keep it useful. A good llms.txt file should explain what the site does and point agents to important pages. For Layzr, that connects directly to AI brand visibility: agents need a clean version of your brand story before they can repeat it accurately.

How to check your Lighthouse agentic browsing score

Because this category is still rolling out, you may not see it in every Lighthouse surface yet. DebugBear reported that PageSpeed Insights and Chrome DevTools were still using older Lighthouse versions in mid-May 2026. If your report does not show Agentic Browsing yet, use the latest Lighthouse CLI or a tool that exposes the new category.

  1. Install the latest Lighthouse CLI with `npm install -g lighthouse@latest`.
  2. Run `lighthouse --view https://example.com/` against a page you care about.
  3. Open the report and look for the Agentic Browsing category.
  4. Read the fractional pass ratio first, then open each audit.
  5. Separate informational checks from problems that would break a real agent task.

Where Layzr goes further

Lighthouse is good at deterministic checks. It can tell you if a label is missing, if CLS is bad, or if a WebMCP annotation is invalid. It cannot tell you whether the page explains the product well, whether the next step feels obvious, or whether an agent-style task actually reaches the right outcome.

Layzr is built for that wider audit. Use it with the new Lighthouse category when you need to check agent-style task behavior, AI visibility, AI brand mentions, design clarity, copy, UX, SEO, and conversion flow.

  • Lighthouse agentic browsing checks machine-readable structure.
  • Layzr checks whether the page is understandable, persuasive, and usable.
  • Lighthouse looks at one URL from a technical angle.
  • Layzr can connect the page audit with AEO, AI visibility, brand mentions, and Search Console context.

Agent-friendly website checklist

This is the practical version. If you want to prepare for agent-friendly websites, start with the boring fixes. They help agents, but they also help people.

  • Use semantic HTML for buttons, links, forms, navigation, and headings.
  • Give every meaningful form control a label.
  • Make important states visible and programmatic.
  • Reserve space for images, embeds, ads, and late-loading UI.
  • Avoid transparent overlays and ghost elements that cover interactive controls.
  • Keep targets large enough to detect and click.
  • Add llms.txt only if it gives agents a useful summary and real links.
  • Use WebMCP carefully when you have forms or actions agents should understand.
  • Run PageSpeed Insights for CLS and loading issues.
  • Run Layzr to catch the design, copy, SEO, and agent-task problems Lighthouse misses.

How agentic browsing connects to AEO

Agentic browsing and answer engine optimization are not the same thing, but they are getting closer. AEO asks whether AI systems mention and describe your brand correctly. Agentic browsing asks whether an AI agent can read and act on the page once it arrives.

The overlap is obvious: clear content, stable layouts, semantic HTML, useful summaries, and visible proof all make the page easier for machines to understand. That is also why Google Search Console still matters. You need to know which pages already have search demand before you decide which pages deserve agent-ready cleanup first.

References

FAQ

What is Lighthouse agentic browsing?

Lighthouse agentic browsing is an experimental Lighthouse category that checks whether a page is constructed for machine interaction. It looks at signals like WebMCP, accessibility tree quality, layout stability, and llms.txt.

How is the agentic browsing score calculated?

Chrome says the category does not use a weighted 0 to 100 score. It shows a fractional pass ratio, audit pass or fail status, warnings, and informational counts.

Does PageSpeed Insights show agentic browsing?

The feature is still rolling out across Lighthouse surfaces. If PageSpeed Insights does not show Agentic Browsing yet, run the latest Lighthouse CLI or use a tool that exposes Lighthouse 13.3 or newer.

What is WebMCP in Lighthouse?

WebMCP is a proposed way for websites to expose forms and actions to AI agents. Lighthouse can detect registered WebMCP tools and check whether forms have declarative metadata like toolname and tooldescription.

Does llms.txt affect the agentic browsing score?

The llms.txt audit is part of Lighthouse's agentic browsing discoverability checks. Chrome says a missing llms.txt file is marked N/A for now because the file is optional, while server errors can be flagged.

How is Layzr different from Lighthouse agentic browsing?

Lighthouse checks deterministic technical signals. Layzr reviews the wider page experience: agent-style tasks, AI visibility, brand mentions, design, copy, UX, SEO, and conversion problems.

Test the page beyond Lighthouse

Use Layzr to review agent-style task behavior, AI visibility, brand mentions, UX, copy, SEO, and the parts Lighthouse cannot judge from a deterministic audit.

Run an agent-style test