Lighthouse Now Audits Your Site for AI Agents. Here's How to Pass the Test.

TL;DR

• Lighthouse's new Agentic Browsing category runs deterministic checks for WebMCP integration, an accessibility tree built for machine reading, layout stability, and discoverability via llms.txt.
• It does not produce a 0–100 score. Instead, you get a fractional pass ratio — Google's signal that the standards are still emerging and they want data, not a rank.
• The audit reuses signals you already know — semantic HTML, ARIA labels, low CLS — and adds two new ones most teams have never heard of: WebMCP (a way to expose your forms and actions to agents) and llms.txt (a Markdown summary of your site for LLMs).
• Google's own SEO team told marketers to ignore llms.txt. Google's Chrome team now audits for it. That gap is the story.
• The five fixes below get most B2B SaaS sites from "invisible to agents" to "actionable" in under two sprints.

What is the Lighthouse AI agent audit?

The Lighthouse AI agent audit is an experimental category inside Chrome's Lighthouse tool, called Agentic Browsing, that evaluates how well a website is built for machine interaction by AI agents. It checks four areas: WebMCP integration (the protocol that lets agents call your site's functions), an accessibility tree that agents can parse, layout stability that lets agents reliably click the right element, and discoverability via llms.txt. Unlike Performance or SEO, it reports a pass ratio instead of a 0–100 score.

The audit lives in Chrome DevTools' Lighthouse panel alongside the four classic categories. Your coding agent can also trigger it through DevTools for agents and act on the findings directly — a workflow Google now calls "agentic browsing" rather than "AI SEO."

Why this matters now

For the past two years, the conversation about AI search has been about citations: are you getting quoted inside an AI Overview, a ChatGPT answer, a Perplexity response. That's still the visibility layer. But the layer underneath — agentic readiness — is moving fast, and Lighthouse is the first widely deployed tool that scores it.

Three forces are converging in 2026. First, agent traffic is real: Cloudflare's 2025 Radar data showed AI-related bot requests grew faster than any other crawler category through the year. Second, browsers are shipping agent rails. Chrome's DevTools team now treats agentic browsing as a first-class use case, not a curiosity. Third, the WebMCP proposal — a standard way for a page to advertise the actions an agent can take — is in active developer preview behind a flag.

Translation for B2B marketing leaders: the same way mobile-first indexing forced a rebuild of how sites were structured between 2015 and 2018, agentic-first indexing is the next forcing function. The teams that catch this early get the early-mover citation and conversion advantage on the agent side. The teams that ignore it spend 2027 doing emergency rewrites.

What Lighthouse checks (and how it scores)

The Agentic Browsing category runs a fixed set of deterministic audits and reports a pass ratio — for example, 6/9 — not a percentage. Google has said the standards are emerging, so the report is designed to surface actionable signals rather than rank sites against each other. Audits can return Pass, Fail, or Not Applicable. Errors fire only when a technical requirement is violated (for example, an invalid WebMCP schema). Results can vary between runs if your tools register on a delay, your DOM shifts during the snapshot, or layout instability moves elements.

Here's the audit map in one table.

Audit Category What it checks Pass condition Registered WebMCP tools WebMCP integration Lists all WebMCP tools registered on the page via the Declarative or Imperative API. Informational — at least one tool registered. WebMCP schema validity WebMCP integration Validates the schema of each registered tool. No schema errors. Forms missing declarative WebMCP WebMCP integration Detects <form> elements that should expose a WebMCP tool but don't. No flagged forms. Accessibility for agents Accessibility Subset of a11y checks critical for machine reading: names/labels, tree integrity, visibility of interactive elements. All required a11y rules pass on interactive elements. Cumulative Layout Shift Layout stability Measures unexpected layout shifts that move elements between agent identification and interaction. CLS within Lighthouse's stability threshold. llms.txt Discoverability Checks for https://yourdomain.com/llms.txt. File served, or marked Not Applicable.

The category header surfaces a pass ratio at a glance. Individual audits can fire errors or warnings that your dev team's agent can act on directly inside DevTools.

How the audit reframes "AI SEO" for marketing leaders

For the past two years, AI SEO advice has split between two camps. One camp told you to chase citations by writing extractable answer blocks. The other told you to chase rankings by doing the same SEO you've always done, just sharper. Lighthouse's new category quietly adds a third layer: interaction readiness.

The new question is no longer "can an LLM read my page?" It's "can an agent acting on behalf of my buyer complete the next step on my page?"

That changes which pages matter. The blog post that gets cited still matters for awareness. The pricing page, the demo request form, the comparison table, the checkout flow — those become the surfaces that get scored for whether they can be operated by an agent. If they can't, the agent will route the buyer to a competitor that can.

The 5 fixes that get most B2B SaaS sites to "agent-ready"

These are the actions that move the Lighthouse Agentic Browsing pass ratio fastest, ranked roughly by leverage per hour of engineering time.

1. Run the audit on your top 10 revenue pages. Open Chrome DevTools, navigate to Lighthouse, enable the Agentic Browsing category, and run it against your homepage, pricing page, demo form, top three feature pages, top three blog posts, and your highest-converting landing page. Note the pass ratio per page. This is your baseline.
2. Fix the accessibility tree on every interactive element. Real <button> elements with aria-label where text is ambiguous. Real <input type="email"> instead of styled <div>s. Logical heading hierarchy. Every clickable thing needs a programmatic name. This work also lifts your normal Lighthouse Accessibility score and your WCAG posture — three checkboxes, one effort.
3. Add declarative WebMCP to your high-intent forms. For each form an agent might fill — demo request, contact, sign-up, pricing inquiry — add toolname and tooldescription attributes to the <form> element. Use action-oriented names like toolname="request_demo" and descriptions that explain when an agent should use it ("Use this tool to request a sales demo for a B2B SaaS team of 10+ people"). This is the highest-leverage new fix on the list.
4. Kill CLS on the routes agents will use. Image dimensions, font loading strategy, no late-injected banners on revenue pages. Layout shifts move buttons between the moment the agent identifies them and the moment it tries to click. Agents misfire silently — you won't see a bug report.
5. Ship an llms.txt file at your root. A short Markdown file at https://yourdomain.com/llms.txt with a one-paragraph summary of what your site does and links to your most important pages (pricing, docs, demo, case studies). Google's Search team says it isn't a ranking signal for Google AI features. Anthropic and Mistral have publicly used llms.txt files in their own developer docs, and Lighthouse audits for it. Cost is one hour. Downside is zero.

Common mistakes when preparing a site for AI agents

These are the patterns we see most often when auditing B2B SaaS sites against the new Lighthouse category.

• Treating it as a developer problem. It isn't. The forms that matter for revenue — demo, pricing, contact — sit in your marketing site, owned by the team that runs your CMS. Marketing leaders need to own the agentic readiness roadmap, not delegate it to engineering.
• Optimizing only the homepage. Agents land where the answer points them, which is rarely your homepage. The Lighthouse audit needs to be run page by page on the routes that drive pipeline.
• Confusing llms.txt with robots.txt. robots.txt controls who can crawl. llms.txt is a summary for LLMs. They solve different problems and don't replace each other.
• Skipping WebMCP because it's experimental. It is experimental. It's also the only standardized way to tell an agent what it can do on your site. Early adopters get an outsized share of the first wave of agent traffic, the same way mobile-first sites did in 2016.
• Assuming Google's word is final. Google's Search team has publicly said llms.txt isn't used by Google's AI features. Google's Chrome team built an audit for it anyway, because ChatGPT, Claude, Perplexity, and Copilot are not Google. Your buyers don't only use Google.

Most B2B teams in 2026 are still operating on layer one. Layer two — the citation gap — is where competitors are quietly winning new pipeline. Layer three is the surface the rest of the decade will be built on. The Lighthouse audit is the first deterministic, reproducible signal at that bottom layer. Use it.

Frequently Asked Questions

What is the Lighthouse Agentic Browsing audit? It's a new, experimental Lighthouse category in Chrome DevTools that evaluates how well your website supports AI agents. It runs deterministic checks for WebMCP integration, accessibility tree quality, layout stability, and llms.txt discoverability, and reports a pass ratio instead of a 0–100 score.

How is it scored? The Agentic Browsing category does not produce a weighted 0–100 score. Lighthouse reports a fractional pass ratio (for example, 6/9) showing how many agentic readiness checks your site passes, plus pass/fail status and warnings on individual audits. Google chose this format because the underlying standards are still emerging.

Do I need an llms.txt file? You don't strictly need one. The Lighthouse audit marks llms.txt as Not Applicable if the file is absent, not as a failure. Google's Search team has said it isn't a Google ranking signal. But ChatGPT, Claude, Perplexity, and other AI agents can use it, and the file costs an hour to ship — so most B2B sites should add one.

What is WebMCP and do I need to implement it? WebMCP is an experimental web standard that lets your page explicitly expose actions — "book a demo," "add to cart," "request a quote" — to AI agents. You declare tools by adding toolname and tooldescription attributes to <form> elements, or by registering them in JavaScript. It's the highest-leverage new fix in the Lighthouse Agentic Browsing audit for B2B SaaS forms.

Is this the same thing as AI SEO or GEO? No. AI SEO and Generative Engine Optimization (GEO) focus on whether your content gets cited inside AI answers. The Lighthouse Agentic Browsing audit focuses on whether an AI agent can navigate and act on your pages after the buyer arrives. They're complementary layers — both matter, neither replaces the other.

Will this audit affect my Google ranking? Not directly. Google has been explicit that the Agentic Browsing category is experimental and gathers data rather than ranking sites. But the underlying signals — accessibility tree quality, low CLS, semantic HTML — overlap heavily with classic SEO and Core Web Vitals, so passing the audit tends to lift your traditional rankings as a side effect.

The next decade of growth is being decided in the agentic browsing layer

Lighthouse just gave marketers something they've never had before: a deterministic, reproducible scorecard for whether AI agents can use the website you're paying to build. The tactical work — semantic HTML, ARIA labels, low CLS, llms.txt, declarative WebMCP — is concrete, cheap, and mostly done by your existing engineering team. The strategic shift is harder: stop treating AI as a content problem and start treating it as an interaction problem. Your buyers are about to send agents instead of clicks. Run an OutAnswer AI visibility audit to see where your brand is being cited today, and which agent-readable surfaces your competitors have already shipped.

Related Reading

• Google Says AEO and GEO Are 'Still SEO.' Here's What That Actually Means for B2B CMOs. — How Google's official AI search guide changes the playbook
• Agentic Browsing Audits — Chrome for Developers — Google's official scoring documentation
• WebMCP API documentation — How to register WebMCP tools on your site
• llms.txt specification — The emerging standard for machine-readable site summaries