Google Lighthouse now audits whether AI agents can find your site’s tools

Google’s Lighthouse 13.5 release adds an audit for Agentic Resource Discovery, a proposed specification for how AI agents locate the tools and services a website offers. It is a technical, forward-looking addition, and it signals where Google expects site auditing to head next.

What the audit checks

Lighthouse checks three locations, in order, for a pointer to a site’s ARD catalog:

• An Agentmap directive inside robots.txt
• A link tag carrying the ai-catalog relation
• The same relation inside the HTTP Link header

If none of the three exists, Lighthouse falls back to requesting /.well-known/ai-catalog.json directly. When no pointer and no file are found, the audit is marked Not Applicable rather than failed. A schema error, or a pointer that leads to a catalog Lighthouse cannot load, does count as a failure.

The new check sits inside Lighthouse’s experimental Agentic Browsing category, grouped alongside the existing llms.txt audit under a heading called Agent Discoverability. This category does not produce a 0 to 100 score. Google’s documentation attributes that choice to the standards for the agentic web still being in early development. The update is expected in Chrome 156 DevTools, with a PageSpeed Insights rollout to follow within two weeks.

A specification that has already moved on

The timing here matters. The ARD specification has been updated since Lighthouse 13.5 shipped its detection logic.

Version 0.91 of the spec now names /.well-known/ard.json as the manifest location, with the earlier Ai-catalog.json path kept only as a legacy reference. The spec itself notes that a file sitting only at the old path may not be found by newer readers. Lighthouse 13.5’s source code, as of its release, does not check for Ard.json or an ard link relation anywhere.

Practically, this means a site correctly following the newest version of the spec could still show as Not Applicable in Lighthouse today. That result reflects what Lighthouse’s current detection logic looks for, not necessarily whether a valid catalog exists on the site.

Where ARD fits next to llms.txt and WebMCP

These three specifications solve different problems, and the distinction is easy to blur.

• ARD helps an AI agent discover callable services: MCP tools, A2A agents, and skills a site exposes.
• llms.txt offers agents a plain-language summary of a site’s content.
• WebMCP lets a page expose structured actions an agent can call once it has already arrived on that page.

None of this is tied to Google Search ranking. The release notes make no connection between the Agentic Browsing category and search visibility, and Google’s own guidance elsewhere has already drawn a line between how Search treats a site and how Lighthouse’s agent-focused checks evaluate it.

What this signals

The specification carries backing from authors at Google, Microsoft, and Hugging Face, and Lighthouse has now built tooling around it before the spec has fully stabilised. That sequence, tooling arriving ahead of a settled standard, is becoming a familiar pattern across the agentic web, and it is worth tracking even while the underlying format is still moving.

For now, an ARD catalog is not something most sites need to prioritise over core technical SEO or AI Overview readiness. It is an early signal of a discovery layer built specifically for autonomous agents, distinct from how search engines or AI Overviews find and cite a page, and one that will be worth revisiting once the specification settles on a stable file path.