AEO
Contents
On June 17, 2026, Google and a broad industry coalition published an open specification called Agentic Resource Discovery, or ARD. It gives AI agents a standard way to find a business’s tools and services and confirm who owns them. The spec is a genuine piece of work with serious companies behind it. It’s also something almost no one needs to act on right now, and the reason why tells you more about AI visibility than the spec itself does.
What is Google’s ARD spec?
ARD is an open specification for publishing, discovering, and verifying AI capabilities across the web. A company uses it to advertise its callable resources — tools, APIs, skills, and agents — in a standard format, so other AI agents can find those resources, check who really owns them, and connect. Google released it under the Apache 2.0 license at version 0.9, which means it’s a draft proposal, not a settled standard.
Google calls ARD “the missing layer of the agentic web.” The operative word is layer. ARD handles discovery: it helps an agent figure out where a capability lives and whether to trust it, then steps aside.
How does ARD actually work?
Two pieces do the heavy lifting: catalogs and registries.
The ai-catalog.json file
A business publishes one static file, ai-catalog.json, listing the capabilities it wants agents to find — an MCP server, an agent, an API, or a dataset. Each entry says what the capability is, where it lives, and how to verify it. It’s the agent-era cousin of sitemap.xml, except it indexes things an agent can do rather than pages a person can read.
The /.well-known/ path and domain-as-identity
The file typically sits at a fixed location on your own domain: https://yourdomain.com/.well-known/ai-catalog.json. That placement is the core of the trust model. A catalog can only live on the domain it claims to represent, so an agent reading the catalog at acme.com knows those capabilities belong to Acme. Nobody can publish a catalog impersonating your bank from a domain they don’t control.
In addition to the well-known path, the ARD specification permits discovery through two other primary methods: an Agentmap directive in your robots.txt file (e.g., Agentmap: https://yourdomain.com/custom-path/ai-catalog.json) or a <link rel="ai-catalog" href="..."> tag in your HTML head.
Registries — search engines for the agentic web
Registries crawl these published catalogs, index them, and answer plain-language queries from agents: “find a tool that checks shipping rates.” They return ranked, verifiable matches. The agent picks one, ARD passes along the trust details, and the agent connects directly using that capability’s own protocol.
One line captures the role: MCP solved interaction; ARD solves discovery. MCP (Model Context Protocol) and Google’s Agent2Agent protocol define how agents talk to tools and to each other. ARD sits in front of them and answers: which tool, whose is it, and can it be trusted?
Who is behind ARD, and who isn’t?
The launch coalition is broad: Google, Microsoft, GitHub, Hugging Face, GoDaddy, Cisco, Databricks, Nvidia, Salesforce, ServiceNow, and Snowflake, built on an AI Catalog data model governed by a Linux Foundation working group. This is an open, multi-vendor standard, not a Google-only product.
OpenAI isn’t on the list. The company behind ChatGPT, the most-used consumer AI assistant, has not signed on to the discovery layer Google is proposing. Notably, OpenAI is a founding member of the Linux Foundation’s newly established Agentic AI Foundation (alongside Anthropic and Block) which stewards related standards like AGENTS.md and MCP, but it has steered clear of this specific ARD draft. That doesn’t kill ARD, but it’s a fair reason to be skeptical of anyone calling this the settled future of AI discovery.
Is ARD live yet?
Barely. A June 2026 adoption census of 39 major sites — including companies that helped write the spec — found not a single one serving a working ai-catalog.json. The spec is still a v0.9 draft. Google’s own implementation, an Agent Registry in its Gemini Enterprise Agent Platform, lists native support as months out.
So: real technology, serious backers, near-zero deployment. That’s the case for paying attention without changing anything yet.
Does ARD help my content get cited by AI?
No, and this is where most people misread it. ARD helps agents discover callable capabilities — tools, APIs, and bookable services. It does nothing to make your blog post or service page get quoted by ChatGPT, Gemini, or Google’s AI Overviews.
Getting cited by an AI answer and being callable by an AI agent are two different layers of the same shift, and most businesses haven’t won the first one yet. If your company is content plus a contact form, you have nothing to put in a catalog. There are no callable capabilities to list.
ARD, llms.txt, Schema.org, AGENTS.md: which matters now?
Several “publish a file so AI understands you” conventions are floating around. They’re layers of a stack, not rivals. Here’s how they rank for a typical business today.
| Standard | What it exposes | Priority for most SMBs today |
|---|---|---|
| Schema.org structured data | What your business and pages are (organization, products, services, FAQs) | High — foundational for search engines and AI engines alike |
| llms.txt | A markdown guide pointing AI models to your key content | Medium — cheap, lightweight, and useful now |
| AGENTS.md | Instructions and boundaries for coding agents working within a codebase | Situational — mainly developer-facing organizations |
| ARD (ai-catalog.json) | Your callable tools, APIs, and agentic capabilities | Low today — unless you build or sell callable capabilities |
The work that pays off now sits in the top rows. Clean structured data and a clear content footprint help AI systems understand and recommend you today, whether or not ARD ever catches on.
What should you do today?
If you sell tools, APIs, or agents
You’re the exception that should pilot ARD now. If you expose booking, availability, quoting, e-commerce, or an actual API, publishing a small, well-formed ai-catalog.json is a reasonable early bet. Keep it scoped. It’s a pilot, not a platform.
If you’re a content or service business
Skip the catalog. Fix the layer that decides whether AI knows you exist: consistent business data across the web, real Schema.org markup, an llms.txt file, and content that answers the questions buyers actually ask. None of that depends on ARD winning, and all of it helps you today.
If you’re an agency
ARD is a topic your clients will raise within a year, and few agencies can discuss it accurately. Brief them straight: here’s the layer, here’s why it can wait, and here’s what we’re doing instead. Then offer discovery-layer monitoring as an ongoing service. It’s a retainer you can run, or have run under your brand, without overpromising.
The bottom line
- ARD is a real, well-backed standard for how AI agents discover and trust business capabilities.
- It covers tools, not content, so it won’t get your articles cited by AI. That’s a separate problem.
- Adoption is near zero, even among its authors, and OpenAI isn’t involved in the draft. Watch it; don’t rush it.
- The work worth doing now is the foundation: schema, consistent business data, and
llms.txt.
What does AI actually see when it looks at your business?
Before worrying about the discovery layer, find out whether today’s AI even mentions you. We’ll run your site through a free AI-visibility check: which engines cite you, which cite your competitor, and the three fixes that would move the needle. No pitch, just the results. Get your free AI-visibility audit →
Run an agency? We do these audits, and the schema and data work behind them, white-label under your brand.