Google shipped Lighthouse 13.5 with a new experimental audit that looks for how AI agents find a site’s tools and services. The audit, built around a proposal called Agentic Resource Discovery (ARD), checks three locations for a catalog pointer before falling back to a default file path. That fallback path is already out of date against the spec it was built to validate.

The stakes here are narrow but real. Sites that want AI agents such as MCP clients or A2A systems to discover their callable services now have a Lighthouse signal to check against, separate from anything related to organic ranking. Search Engine Journal’s Matt G. Southern, who reviewed the 13.5 source code, found that Lighthouse still searches for the older ai-catalog.json filename even though the ARD project moved to ard.json on Aug. 26.

Lighthouse checks a site’s robots.txt file for an Agentmap directive, a link tag carrying the ai-catalog relation, and the HTTP Link header carrying that relation too. When all three checks come up empty, the audit requests /.well-known/ai-catalog.json directly. A schema mismatch fails the audit; an unreachable catalog behind a valid pointer fails it too. Southern reported finding no reference to ard.json or an ard link relation anywhere in the 13.5 source or the project’s main branch as of Sept. 21.

That gap matters because ARD version 0.91 requires compliant software to fetch ard.json first, with the older path optional. The spec itself warns that a manifest published only at the old location “may not be found.” A site that migrated fully to the new filename and dropped the legacy path, exactly as the spec anticipates, would register as Not Applicable in Lighthouse rather than pass. Not Applicable does not mean a site lacks a current ARD manifest; it means Lighthouse’s current search areas did not locate one.

Google’s own documentation acknowledges the instability directly, describing the Agentic Browsing category as one that skips a 0-100 score in favor of a pass ratio because “the standards for the agentic web are still emerging.” The 13.5 release also pairs the new ARD check with the existing llms.txt audit inside a shared “Agent Discoverability” heading, and it introduces an automated weekly job watching for upstream changes to the ARD schema or conformance tests, an acknowledgment that the target keeps moving.

This is worth separating clearly from Search guidance. The release notes never connect ARD to Google Search, and Southern’s reporting draws the same line he drew when Lighthouse added its llms.txt audit in May: Search’s ranking guidance and Lighthouse’s agent checks evaluate different things. Nothing here is evidence of a ranking factor, and this audit category carries no signal about traffic or visibility, experimental or otherwise.

The practical distinction that matters for site owners: ARD is built for machine-callable services (MCP tools, A2A agents, skills), where an llms.txt file summarizes page content for agents and WebMCP lets a page expose actions an agent can trigger once it has already arrived. The three are complementary discovery layers, not competing ones.

Teams already publishing an ARD manifest should keep the legacy ai-catalog.json path live alongside the new ard.json file until Lighthouse’s checks catch up, since the spec’s own authors, drawn from Google, Microsoft, and Hugging Face, built the fallback allowance for exactly this transition period. Treat a Not Applicable result in PageSpeed Insights over the next two weeks as inconclusive, not as a fix-it signal, until Lighthouse’s gatherer code is updated to match the current spec.

Search Engine Journal (Matt G. Southern) reported this Sept. 21, 2026.