On June 17, 2026, Google published the Agentic Resource Discovery (ARD) specification alongside ten industry partners: Cisco, Databricks, GitHub, GoDaddy, Hugging Face, Microsoft, NVIDIA, Salesforce, ServiceNow, and Snowflake. The spec defines a standard for how AI agents discover, evaluate, and connect to capabilities across the web.
You probably haven’t implemented it yet. Almost no one has. That’s exactly the problem — and the opportunity.
ARD is not another rebranding of SEO. It’s a new protocol that answers a different set of questions than search engines ask. And if you run a website that offers services, tools, or information that AI agents might need, implementing it today puts you in the 1% that’s discoverable to the next generation of web traffic before that traffic becomes competitive.
What Problem ARD Solves
Search engines answer human queries. AI agents need to answer a harder set of runtime questions before they can act on your behalf: where does the right capability live, which provider should I use, and how do I verify it’s safe to connect to?
These three questions — location, selection, and verification — are what the ARD spec is designed to answer at machine speed. Humans resolve them through browsing, reading reviews, and making judgment calls. AI agents need structured, machine-readable signals they can parse without human help.
The current state of the web is that AI agents largely can’t do this. They rely on training data (which is frozen), llms.txt files (which have no registry or query mechanism), and AGENTS.md hints (which are informal). ARD introduces two formal primitives that change this: a static ai-catalog.json manifest and a registry API that indexes those manifests and returns ranked results to natural-language discovery queries.
The Two Primitives: ai-catalog.json and the Registry
The ai-catalog.json file is hosted directly under your domain — typically at https://yourdomain.com/ai-catalog.json or at https://yourdomain.com/.well-known/ai-catalog.json. Because it’s served from your domain, ownership of that domain acts as the cryptographic foundation for identity and trust. An AI agent querying the registry can verify that the capability being offered actually belongs to the domain it claims.
The registry is a crawl-and-index layer, similar in concept to a search engine index but purpose-built for agent discovery queries. GitHub Copilot’s new Agent Finder is the first publicly visible implementation — it uses ARD-indexed catalogs to surface capabilities when developers ask questions like “find me a tool that can review my pull requests.”
The spec is licensed under Apache 2.0 and builds on the AI Catalog data model from the Linux Foundation’s AI Catalog Working Group, which means it’s designed to be open infrastructure, not a proprietary Google lock-in.
What Goes Inside ai-catalog.json
The AI Catalog data model defines several key fields. A minimal implementation for a service business might look like this:
{
"@context": "https://ai-catalog.org/schema/v1",
"name": "YourBrand AI Services",
"description": "SEO and GEO optimization services for WordPress sites",
"provider": {
"name": "Your Company Name",
"url": "https://yourdomain.com",
"contact": "hello@yourdomain.com"
},
"capabilities": [
{
"id": "geo-audit",
"name": "AI Visibility Audit",
"description": "Scans a website and scores its visibility across ChatGPT, Perplexity, and Google AI Overviews",
"type": "tool",
"endpoint": "https://yourdomain.com/api/v1/audit",
"auth": "none",
"pricing": "free"
}
],
"tags": ["seo", "geo", "ai-visibility", "wordpress"],
"last_updated": "2026-08-09"
}The capabilities array is where the substance lives. Each capability describes a specific thing your site can do for an AI agent — a tool it can call, a resource it can retrieve, a service it can book. The more precisely you describe each capability, the more accurately the registry can match your site to agent queries.
For informational sites and publishers (rather than tool providers), the capability type shifts from tool to resource, and you’d describe the types of information you publish, the update frequency, and the access method (RSS, API, or crawl-friendly HTML).
How GitHub Agent Finder Uses ARD Right Now
GitHub Copilot’s Agent Finder is the most concrete live example of ARD in production. When a developer asks Copilot to “find a code review tool” or “find a deployment automation agent,” Copilot queries the ARD registry and surfaces ranked results from indexed ai-catalog.json files.
The ranking criteria in Agent Finder mirror what you’d expect from a well-designed discovery system: specificity of the capability description, recency of the last_updated field, verified domain ownership, and match quality between the capability description and the natural-language query.
This is the earliest signal of what AI agent discovery will look like across the broader web within 12-18 months, as other AI systems integrate ARD-compatible registries. Being indexed now, while competition for registry placements is minimal, is structurally similar to buying Google Ads in 2002 or getting indexed by Bing in 2009 — the first movers establish positions before auction pressure or competition changes the economics.
ARD vs. llms.txt vs. AGENTS.md: Where Each Fits
These three files serve different layers of the agent-readiness stack and are not substitutes for each other:
- llms.txt — hints to AI crawlers about what content to prioritize when training or retrieving. Passive signal, no registration, no query mechanism. Scope: content retrieval.
- AGENTS.md — instructions to AI agents about how to behave on your site: what they can do, what’s off-limits, how to authenticate. Scope: behavioral rules for site interactions.
- ai-catalog.json (ARD) — active registration of your capabilities in a discoverable registry. Scope: getting found by agents that are actively searching for what you offer.
A complete agent-ready implementation uses all three. But if you only have bandwidth for one new file, ai-catalog.json has the highest leverage in 2026 because it’s the only one that makes you findable, not just readable.
How to Implement ARD on WordPress in 20 Minutes
WordPress makes this straightforward because you can serve a static JSON file at any URL. Here are the steps:
Step 1 — Create your ai-catalog.json file
Use the schema above as a template. Fill in your actual services, tools, or information resources as capabilities. Be specific: “SEO content optimization” is more discoverable than “marketing services.”
Step 2 — Upload to your WordPress root
Upload the file to your server’s root directory (the same level as wp-config.php) so it’s accessible at https://yourdomain.com/ai-catalog.json. You can do this via FTP, your hosting file manager, or the SFTP terminal.
Step 3 — Verify it’s serving correctly
Visit https://yourdomain.com/ai-catalog.json in your browser. You should see raw JSON. If you get a 404, you may need to add a rewrite rule to .htaccess:
RewriteRule ^ai-catalog\.json$ /ai-catalog.json [L]Step 4 — Add the well-known path (recommended)
The ARD spec supports both root-level and well-known paths. Add the file at /.well-known/ai-catalog.json as well, since some agents may query that path specifically. Create a .well-known folder in your root and place a copy there.
Step 5 — Register with emerging ARD registries
GitHub’s Agent Finder is the first public registry. Submit your ai-catalog.json URL to GitHub’s registry when the submission form becomes publicly available (currently in closed beta). Watch for similar registry submissions from Google and Microsoft in Q3-Q4 2026.
Step 6 — Add a meta tag pointing to the catalog
In your WordPress theme’s header.php, add this line to the <head> section so AI crawlers that parse HTML can discover your catalog:
<link rel="ai-catalog" href="https://yourdomain.com/ai-catalog.json" />The Window Is Open, But Not Indefinitely
Public adoption of the ARD spec is currently very low. The Synscribe analysis of the spec’s first 60 days found that fewer than 3,000 domains had published a valid ai-catalog.json file — out of an estimated 1.9 billion active websites. That ratio is a feature, not a bug, if you move now.
The pattern from llms.txt is instructive: when it launched in late 2024, early adopters saw 37% higher AI crawler request rates than late adopters measured in 2026 studies. ARD is at an earlier stage of adoption, with a more formal backing consortium, which suggests the adoption curve will be steeper and the early-mover advantage more durable.
The three-step minimum for 2026: publish llms.txt (content retrieval), publish AGENTS.md (behavioral rules), publish ai-catalog.json (active discoverability). All three are static files. None require a developer. All three together represent the agent-readiness baseline that forward-looking sites are implementing before AI agents become a primary traffic source.
Check where your site currently stands with AI agents — including whether you appear in ChatGPT, Perplexity, and Google AI Overview answers — at ai-visibility.llmagnet.com. The audit runs in 60 seconds and identifies the specific gaps.