Lighthouse’s “agentic browsing” audit: what it checks, how to pass

For fifteen years, Lighthouse categories have defined what "a good website" means in practice: performance, accessibility, best practices, SEO. That list now includes agentic browsing — Lighthouse audits whether AI agents can operate your site, with checks built around WebMCP. When Lighthouse measures something, budgets and dashboards follow; agent readiness just became a line item.

What the checks look for

The agentic-browsing category currently centers on three questions:

1. Does the page expose WebMCP tools? Lighthouse inspects the rendered page for tools registered via document.modelContext (the imperative API) or declarative form-tool annotations. No tools, no points — a page that's merely crawlable is readable, not operable.
2. Are the tools well-formed? Names present and within limits, descriptions non-empty (they're prompts for the model — quality matters), and inputSchema parseable as JSON Schema. A tool with a vague description is flagged the way an image without alt text is: technically present, practically useless.
3. Is the environment sane? Secure context (HTTPS), no document.domain origin games, and — during the origin-trial window — a valid trial token so the API actually exists for real Chrome users.

How to run it

# Chrome DevTools → Lighthouse tab → check "Agentic browsing"
# or CLI:
npx lighthouse https://yoursite.com --only-categories=agentic-browsing --view

Run it on your money pages, not just the homepage — the audit is per-page, and agents land deep. A store whose product pages score zero while the homepage passes has passed a vanity check.

Passing, step by step

• No tools at all? Start with one read-only tool wrapping your site's core query — search, availability, pricing. Our cookbook has per-industry patterns with code.
• Tools flagged for descriptions? Rewrite them for the model: what the tool does, when to call it, what it returns. "Search" is a bad description; "Search the product catalog; returns names, prices, availability and URLs" passes and performs.
• Schema errors? Validate your inputSchema objects as real JSON Schema — a stray comment or trailing comma silently voids the tool.
• Token issues? Register for the WebMCP origin trial and serve the token as a meta tag or header on every page that registers tools.

The bigger picture

Lighthouse's category, Cloudflare's Agent Readiness Score, and Chrome's agent-ready toolkit are three arms of the same push: making "can agents use this site?" a measured, comparable property of the web. Our free scan gives you the cross-standard view in 15 seconds, and the deep audit turns it into an implementation plan.