Google Lighthouse 13.5 adds a new audit for agent discovery, checking whether your site correctly implements ARD (Agentic Resource Discovery), a proposed standard that tells AI agents which tools, APIs, and services your organization offers. The rollout is expected in Chrome 156 DevTools and in PageSpeed Insights within two weeks, per the official 13.5 release notes on GitHub.
What Does the Agent Discovery Audit Actually Check?
The new ard-schema audit validates your site’s ARD catalog against a schema defined by the ARD project itself. According to Search Engine Journal’s review of the Lighthouse 13.5 source code, Lighthouse checks three locations before falling back to a default path.
- An
Agentmapline inrobots.txt(a directive that points to where your catalog lives) - A
<link rel="ai-catalog">tag in your HTML head - An HTTP
Linkresponse header carrying the sameai-catalogrelation - If none of the above:
/.well-known/ai-catalog.jsonas a fallback
If no pointer is found and the fallback path returns nothing, the result is “Not Applicable,” not a failure. Schema errors or a catalog that cannot be loaded cause an outright failure. The audit is grouped alongside the existing llms.txt audit (a plain-text file that summarizes site content for AI agents) under a heading called “Agent Discoverability.”
Both audits sit in Lighthouse’s experimental Agentic Browsing category, which is separate from SEO audits. The Lighthouse release notes contain no reference to Google Search ranking. The category does not produce a 0-100 score; it shows a pass ratio, because “the standards for the agentic web are still emerging,” per Google’s own documentation.
What Is ARD, and Who Is Behind It?
ARD is an open specification that lets AI agents automatically find callable resources, MCP servers (a protocol standard that connects AI models to external tools), A2A agents (agents that communicate directly with each other), skills, and APIs, without a human needing to manually wire each connection in advance.
The specification lists three authors: Junjie Bu from Google, R.V. Guha from Microsoft, and Shaun Smith from Hugging Face. Microsoft’s ARD announcement notes it was developed in collaboration with Cisco, Databricks, GitHub, GoDaddy, Google, Hugging Face, Nvidia, Salesforce, ServiceNow, and Snowflake.
The most recent version, v0.91 published August 26, moved the canonical manifest location from /.well-known/ai-catalog.json to /.well-known/ard.json. Software reading ARD files is now required to fetch ard.json, while the older path is optional. The spec cautions that a catalog only available at the old path “may not be found.”
One implementation gap to note: as of September 21, the Lighthouse 13.5 source code and the project’s main branch still reference the older ai-catalog.json naming, not ard.json, according to Search Engine Journal’s code review. The release also does not yet implement ARD’s in-page JSON-LD or DNS discovery methods described in the v0.91 spec.
Two reference implementations are already live. GitHub launched “agent finder,” built on ARD, which lets GitHub Copilot dynamically discover and call the right MCP servers and tools at runtime rather than pre-loading everything. Hugging Face’s Discover Tool provides semantic search across thousands of ML applications and MCP servers using ARD, per Microsoft’s announcement.
What This Means for AI-Search Visibility
Here is the plain-English version: search engines helped people find websites. ARD does the same thing for AI agents finding tools. If your business offers something an AI agent could use, a booking system, a data API, a product search, ARD is how the agent would know you exist at all.
Microsoft’s ARD announcement puts the scale problem directly: the ecosystem of available AI tools already numbers in the hundreds of thousands, but any single agent can only use what has been explicitly wired to it. Everything else “may as well not even exist.” Microsoft draws a direct analogy to the early web, when most users only visited sites already bookmarked in their browser, before search engines solved the discovery problem at scale.
Hingewise analysis: The Lighthouse 13.5 audit carries no confirmed Search ranking signal. But the sequencing is worth noting: the llms.txt audit launched in May, and the ARD audit now joins it in the same “Agent Discoverability” grouping. Lighthouse audits have historically preceded broader guidance from Google about what its systems evaluate. That is not a guarantee this time, but it is a reasonable reason to track the pattern.
The more immediate point is that “AI visibility” is splitting into two distinct problems, and most brands are only thinking about one of them. The first is being cited in AI-generated answers when a user asks a question, this is what content optimization, structured data, and llms.txt address. The second is being available as a callable resource when an AI agent is trying to complete a task, this is what ARD addresses. Lighthouse 13.5 now audits both, but they require different work.
Hingewise’s position: the ARD spec is still a proposal, and Lighthouse’s implementation currently trails the current v0.91 specification. Even so, the depth of industry collaboration, ten collaborating organizations including three AI infrastructure leaders, suggests this is infrastructure being built with serious intent, not a side experiment. Brands that offer callable services should treat ARD implementation as a forward-looking technical hygiene task rather than a wait-and-see question. The cost of adding a well-formed catalog file is low; the cost of being invisible to the next generation of AI agents is not.
Before Your Site Gets Audited: What to Check First
- Open Chrome 156 DevTools (or PageSpeed Insights after the update) and run Lighthouse on your site. Check the Agentic Browsing category for the ARD audit result.
- Check whether your site currently serves a file at
/.well-known/ai-catalog.json. If it does, verify the JSON is valid and matches the ARD schema. - Because ARD v0.91 now requires software to check
ard.jsonfirst, also publish the catalog at/.well-known/ard.json, relying only on the older path “may not be found,” per the spec. - Check your
robots.txtfor anAgentmapline and your HTML for a<link rel="ai-catalog">tag. Either is enough for Lighthouse’s current checks. - Identify which services, APIs, or tools your organization offers that an AI agent could potentially invoke, and assess whether those belong in an ARD catalog at all.
- Keep in mind that Lighthouse 13.5 does not yet check for the
ard.jsonpath orrel="ard"link relation from v0.91, so a “Not Applicable” result today does not necessarily mean your ARD implementation is missing.
The clearest next signal to watch is whether Google’s Search documentation or Search Console begins referencing ARD alongside its current agent-browsing guidance. That would mark the step from experimental infrastructure to something with documented implications for how AI systems find and recommend resources.
Sources: Search Engine Journal (Matt G. Southern, September 2026); Lighthouse 13.5 release notes, GitHub; Microsoft ARD announcement, commandline.microsoft.com.
Sources
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.