Lighthouse 13.5 Adds an ARD Audit — and a Compatibility Gap

Lighthouse 13.5.0, released on September 18, 2026, adds the experimental ard-schema audit for Agentic Resource Discovery (ARD). The audit is designed to check whether a site exposes a machine-readable catalog of callable resources, but the inspected implementation still looks for older ai-catalog discovery names while ARD v0.91 defines newer ard names. That creates a compatibility concern for sites that publish only the current ARD conventions.

Lighthouse 13.5 adds an experimental ARD audit

ARD is a proposed standard for helping AI systems discover callable resources, including MCP tools, A2A agents, skills and other services. Lighthouse places the new audit in its experimental Agentic Browsing category alongside related machine-interaction checks.

The audit identifier is ard-schema. It supports both navigation and snapshot modes, so it can run as part of a page navigation audit or against a captured page state.

The ARD specification, version 0.91, was dated August 26, 2026 and remains a proposal. The compatibility concern became public on September 21, 2026, after the Lighthouse 13.5 implementation was compared with the current discovery conventions in that proposal.

What the Lighthouse ARD audit checks

Lighthouse 13.5 follows a specific discovery sequence:

1. It checks robots.txt for an Agentmap directive.
2. It searches the page for an HTML link using rel="ai-catalog".
3. In navigation mode, it checks the HTTP Link header for rel=ai-catalog.
4. If those signals are absent, it falls back to /.well-known/ai-catalog.json.

Once Lighthouse finds an explicit signal or receives a catalog response with HTTP status 200, the ard-schema audit evaluates the catalog. A catalog that cannot be loaded produces a failing result. A loaded catalog with no errors receives an audit score of 1; low-severity warnings reduce that score to 0.9, while schema errors produce a score of 0.

Lighthouse also uses a pinned version of the ARD conformance implementation and runs a weekly CI check for upstream schema or conformance-test changes.

The gap with ARD v0.91

ARD v0.91 uses different current discovery names. Its well-known manifest path is /.well-known/ard.json, and its current HTML relation is rel="ard".

The older ai-catalog path and relation remain optional compatibility mechanisms in ARD v0.91. Software that implements only the current ard.json path and ard relation still follows the proposal’s current discovery model, while Lighthouse 13.5’s inspected gatherer does not search those newer names.

What “Not Applicable” means here

In Lighthouse 13.5, Not Applicable is returned when the audit finds no explicit discovery signal and receives no catalog response with HTTP status 200. The result reflects the signals and fallback path that this Lighthouse version checks.

That logic creates the reported implementation implication: a site using only /.well-known/ard.json and rel="ard" could receive Not Applicable from Lighthouse 13.5, even though those are the current discovery names in ARD v0.91. The implication comes from the relationship between the inspected gatherer and the proposal’s defined mechanisms.

ARD, llms.txt and WebMCP are different pieces

These technologies address different parts of agent interaction:

• ARD describes how agents can discover callable resources such as MCP tools, A2A agents and other services.
• llms.txt provides a structured summary of a site’s content for language models and other agents.
• WebMCP exposes structured actions that an agent can call on a web page.

Lighthouse groups several of these emerging agent-oriented checks under Agentic Browsing, but an ARD audit is not a search-ranking signal. It evaluates discovery and schema behavior for the resources described by the catalog.

Distribution and scoring in Agentic Browsing

The Lighthouse 13.5 release notes said the update was expected to ship in Chrome 156 DevTools and reach PageSpeed Insights within two weeks of the September 18 release. The Agentic Browsing category uses a fractional pass ratio instead of the weighted 0–100 score used by many traditional Lighthouse categories.

That format reports the share of checks that pass, along with individual pass or fail results and informational counts. It fits a category built around emerging agent-web standards, where the discovery mechanisms and conformance rules are still evolving.