ARD Checker

Enter any domain to audit its Agentic Resource Discovery (ARD) setup. The checker fetches the site’s ai-catalog.json, validates it against the official JSON Schema, confirms the four discovery signals AI agents look for, tests whether AI crawlers can actually reach the file, and then grades the result.

Domain (the site to check — e.g. example.com)

What is Agentic Resource Discovery?

ARD is the discovery standard for the agentic web — backed by Google and the Linux Foundation — that lets AI agents find, verify, and call the tools, APIs, and MCP servers a website exposes. A site publishes a small ai-catalog.json file at /.well-known/, and registries crawl it so agents can find your capabilities at runtime. If you are optimizing to be found and used by AI, not just ranked by Google, this is the new surface to get right. I implemented the full spec on this site and documented everything that broke: I implemented Google’s Agentic Resource Discovery spec; here’s what broke.

What the ARD Checker validates

Catalog file — fetches /.well-known/ai-catalog.json and confirms it exists and is valid JSON.
Schema conformance — validates the catalog against the official ai-catalog JSON Schema (Draft 2020-12) and reports field-level errors.
robots.txt Agentmap — checks for the Agentmap: directive that points crawlers to your catalog.
Link header — confirms your homepage advertises the catalog in an HTTP Link: header.
Head link tag — confirms a <link rel="ai-catalog"> is present in your homepage’s <head>.
Crawler reachability — requests the catalog as five user-agents (ClaudeBot, GPTBot, a browser, and two Python clients) to confirm a firewall is not silently blocking crawlers.

Each result is a pass or fail with specifics, rolled up into a letter grade, so you can tell at a glance whether your site is genuinely discoverable or just looks that way.

About the crawler-reachability test

This is the make-or-break check: a catalog can be live and public yet invisible to registries if a web application firewall blocks unfamiliar user-agents. The test runs on this server, so it reliably catches user-agent-based blocks, the most common failure mode. It cannot see IP-reputation blocks that would affect only a crawler on a different network, so treat a pass as necessary rather than as absolute proof.

Building or exposing your own agent tools? See how I make this site’s capabilities callable by agents on the MCP connection hub, or read the full ARD implementation story.