Google's Lighthouse 13.5 introduces an audit for Agentic Resource Discovery (ARD), a proposed specification for how AI agents locate the tools and services an organization offers. The Lighthouse team expects the update to reach Chrome 156 DevTools and PageSpeed Insights within two weeks of the release.

The audit compares a site's ARD catalog against the schema defined by the ARD project. If the site provides no catalog pointer, Lighthouse defaults to /.well-known/ai-catalog.json. The ARD specification dated Aug. 26 moved that file to ard.json, keeping the old filename optional for software that still reads it.

The new check sits in Lighthouse's experimental Agentic Browsing category, separate from SEO audits, and is not tied to Google Search anywhere in the release notes. Reports will show either an ARD result or Not Applicable when no catalog is found.

What does the audit check?

The 13.5 source code reveals a four-step discovery chain:

• Lighthouse checks robots.txt for an Agentmap line, an ARD directive pointing to a catalog.
• It looks for a link tag with the ai-catalog relation.
• It checks the HTTP Link header for the same relation.
• If none of the three exists, it requests /.well-known/ai-catalog.json.

If no pointer exists and the well-known path does not serve the file, the audit returns Not Applicable. Schema errors cause failure, as does a pointer to an unloadable catalog. The pull request describes ard-schema as a "schema conformance audit." The same release groups the new check with the existing llms.txt audit under a heading called Agent Discoverability.

Google's documentation says the Agentic Browsing category does not produce a 0-100 score. It shows a pass ratio instead, because "the standards for the agentic web are still emerging."

Has the specification moved ahead of the audit?

Yes. ARD v0.91 now specifies /.well-known/ard.json as the manifest location and refers to ai-catalog.json as the earlier path. When Search Engine Journal covered ARD's launch in June, the specification was a v0.9 draft with ai-catalog.json as the file for organizations to publish.

In v0.91, software reading these files must fetch ard.json and may also check the older path. The spec cautions that a file available only at the old path "may not be found."

As of Sept. 21, the reporter found no references to ard.json or the ard link relation in the 13.5 source or in the project's main branch. The specification also mentions discovery through in-page JSON-LD and DNS records; the 13.5 code checks neither. Lighthouse built its validator on the ARD project's own conformance tests.

The specification remains a proposal and credits three authors: Junjie Bu from Google, R.V. Guha from Microsoft, and Shaun Smith from Hugging Face.

How does ARD differ from llms.txt and WebMCP?

The three standards address different layers of agent interaction with websites:

• ARD covers resource discovery. It includes MCP tools, A2A agents, skills and other callable services, so AI systems can find them through search.
• llms.txt summarizes a site's content for agents.
• WebMCP, shared publicly in August, lets a page offer structured actions an agent can call once it arrives on the site.

What does this mean for site owners?

Based on the Lighthouse 13.5 source code, a website that uses /.well-known/ard.json with rel="ard", has no Agentmap line and avoids the older ai-catalog names would be classified as Not Applicable. That means Lighthouse simply did not find a catalog in its current search areas — not necessarily that the site lacks a current ARD manifest.

As Search Engine Journal noted when Lighthouse introduced its llms.txt audit in May, Google Search's guidance and Lighthouse's agent checks examine different things.

The release also adds a weekly automated check that flags when the ARD project's schema or conformance tests change upstream. With the specification evolving faster than the audit code, that tracking mechanism will determine how quickly Lighthouse catches up to ard.json.