Agentic Resource Discovery (ARD) is a new open specification, published June 17, 2026, that defines how AI agents find, verify, and connect to tools, APIs, and services across the web. Backed by Google, Microsoft, GitHub, Hugging Face, NVIDIA, Salesforce, and six other companies, ARD introduces a standardized layer where agents can search for capabilities at runtime instead of being pre-wired to specific tools. It is the most significant infrastructure shift for the agentic web since llms.txt, and it may matter more.
The spec is a v0.9 draft. It is not final. But same-day implementations from GitHub and Hugging Face suggest this one has real momentum. If you publish tools, APIs, or AI agents, the implications are direct. If you run a content site, the implications are structural: the web is being re-indexed for machine consumption, and the standards being adopted now will shape which brands agents can find and trust for years.
This article breaks down what ARD does, how it works, what it means for GEO and AI visibility, and what you should actually do about it. No hype. Just the spec, the data, and the implications.
What ARD Solves: The Pre-Wiring Problem
Today, if you want an AI agent to use your API, your MCP server, or your tool, you have two options. You manually integrate it, which means writing custom code for every connection. Or you publish documentation and hope a developer wires it up manually. Neither scales.
ARD solves this by moving discovery to runtime. Instead of hardcoding connections, agents query registries that crawl publisher-hosted catalogs. An agent that needs a payment processor, a translation service, or a data provider can search for one, verify the publisher, and connect directly. No pre-wiring required.
The parallel to early web search is exact. Before search engines, you had to know a URL to visit a website. Before ARD, you have to know an endpoint to use a tool. ARD is the discovery layer that makes agentic capabilities searchable.
The contributors clearly see this analogy. Martin Jeffrey, Founder at Harton Works, called ARD “the sitemap, reborn for capabilities rather than pages.” That framing is accurate, and it understates the scope. Sitemaps told search engines what pages exist. ARD tells agents what a company can do.
How ARD Works: Catalogs and Registries
The spec has two components.
1. Catalogs (Publisher-Side)
An organization publishes an ai-catalog.json file at a well-known path on its domain. This file lists the tools, MCP servers, agents, or APIs the organization makes available. Each entry includes metadata: what the capability does, how to connect, and what protocols it supports.
Because the catalog sits on the publisher’s own domain, ARD uses domain ownership as the baseline trust mechanism. If api.yourcompany.com/.well-known/ai-catalog.json lists a payment tool, the agent knows your company published it because it came from your domain.
For production use, publishers can attach trust metadata. This lets agents and registries verify the publisher’s cryptographic identity before connecting. The spec does not mandate a specific trust framework, but it supports existing PKI and signed-assertion patterns.
2. Registries (Discovery-Side)
Registries crawl catalogs across the web, index them, and answer discovery requests from agents. An agent sends a natural-language query like “I need a tool to convert currencies in real-time” and the registry returns matching capabilities from its index.
This separation matters. Publishers control what they expose. Registries compete on indexing quality, search relevance, and trust verification. Agents choose which registries to trust. The architecture is decentralized by design, avoiding a single point of control.
ARD vs llms.txt: Different Problems, Different Layers
The comparison to llms.txt is inevitable. Both ask you to publish a structured file on your domain for machines to read. Both are open, voluntary, and early in adoption. But they address different layers.
llms.txt tells AI crawlers what content to prioritize when reading your site. It is a content-level instruction. ARD tells agents what capabilities your organization offers. It is a service-level declaration.
The distinction maps to two different questions:
- llms.txt answers: “When an AI reads my website, what should it focus on?”
- ARD answers: “When an AI agent needs a capability I provide, how does it find me?”
New Ahrefs data from June 2026 shows that 97% of llms.txt files received zero requests from AI crawlers. Google’s John Mueller confirmed on Search Off the Record that the file cannot help LLMs differentiate one site from another because it is self-reported. This does not mean llms.txt is useless. It means the bots that generate citations barely read it, while coding agents and training crawlers do. The value is narrow but real.
ARD targets a different bottleneck. The problem ARD solves is not “how does AI read my content?” but “how does AI find my tools?” That problem is more acute as agentic workflows grow. GitHub shipped agent finder for Copilot on day one. Hugging Face shipped a discovery tool. The demand exists today, not hypothetically.
For brands thinking about technical GEO strategy, the practical split is: keep your llms.txt maintained for coding agents, focus your energy on structured data and answer-first content for AI citations, and watch ARD adoption if you publish APIs or tools.
Same-Day Implementations: This One Has Traction
Specifications without implementations are theory. ARD shipped with working tools on day one.
GitHub launched agent finder for Copilot. Users can now search for MCP servers, skills, tools, and agents from a chosen registry. Copilot users control what gets connected, but discovery happens through ARD-compatible registries.
Hugging Face released a Discover Tool that searches skills and MCP servers across ARD services. Given Hugging Face’s position in the AI ecosystem, this gives ARD immediate reach among developers building agentic systems.
Cisco tied the spec to its AGNTCY Agent Directory, an open source project under the Linux Foundation. This anchors ARD in existing governance infrastructure rather than creating a new standards body.
Three implementations from three major platforms on launch day is not typical for a draft spec. It signals that ARD solves a real coordination problem that these companies were already spending engineering resources on. Compare this to llms.txt, which launched as a community proposal with no major platform implementations for months. The adoption curves look different.
The Broader Pattern: Structured Files for Machine Consumption
ARD does not exist in isolation. It is part of a pattern that accelerated sharply in mid-2026.
On June 15, two days before ARD, Google Cloud published the Open Knowledge Format (OKF). OKF is a markdown format for packaging organizational knowledge (datasets, metrics, runbooks) so AI agents can read it. It is at version 0.1.
In April 2026, schema.org additions for AI search contexts went live, giving publishers new ways to mark up content specifically for AI extraction.
The pattern across all three efforts is identical. Each asks you to publish a structured file under your own domain so AI systems can consume it without manual integration. Each is early-stage and voluntary. Each bets that machine-readable web infrastructure will matter more, not less, as agentic adoption grows.
For agentic SEO, this pattern is the foundation. When AI agents evaluate vendors, compare tools, or research capabilities, they will increasingly rely on these structured files rather than scraping landing pages designed for human eyes.
What Brands Should Actually Do
ARD is real, but it is early. Here is a prioritized action plan based on what type of organization you are.
If You Publish APIs, MCP Servers, or AI Tools
Start paying attention now. The spec is at v0.9, and the initial implementations from GitHub and Hugging Face are live. Monitor the ARD GitHub repository for changes. Evaluate whether publishing an ai-catalog.json file makes sense for your platform.
Priority: Medium. Timeline: Q3 2026.
If You Are a B2B SaaS Company
You likely have both content and APIs. Your content strategy should focus on answer-first structure, entity authority, and AI citation optimization. Your API strategy should include monitoring ARD adoption. When enterprise procurement teams dispatch AI agents to evaluate vendors, the agents that can discover your capabilities through registries will find you faster than those scraping your pricing page.
Priority: Low-medium. Timeline: Q4 2026.
If You Are a Content-Only Site
You have no direct action to take on ARD today. The spec targets callable capabilities, not content pages. Your priority remains the same: structured data, answer-first content blocks, FAQ schema, and building entity authority through mentions across trusted domains. Run an AI visibility audit to see where you stand. If AI engines do not mention you now, a catalog file will not fix that.
Priority: Monitor only. Timeline: 2027.
For Everyone
Three things matter regardless of what you publish:
Your domain is your identity. ARD, OKF, and llms.txt all use domain ownership as the baseline trust signal. Own your domain. Host your structured files there. Do not delegate this to a third-party subdomain.
Structured data is compounding. Every structured file you publish (schema markup, llms.txt, future ARD catalogs) adds to a machine-readable layer that AI systems can consume. Start with what exists today. Add new formats as adoption grows.
AI visibility is the prerequisite. If AI engines do not mention your brand today, infrastructure will not save you. The first step is always measuring where you stand. You can check yours in 60 seconds at audit.searchless.ai.
The Risk of Building for Systems That Never Arrive
Google’s John Mueller has been candid about this risk. He argued that LLM systems cannot use files like llms.txt to distinguish one site from another, and he advised focusing on current needs rather than future agent-oriented strategies. The same caution applies to ARD.
The counterargument is the implementation evidence. GitHub, Hugging Face, and Cisco shipped working tools on day one. Google plans ARD support in its Gemini Enterprise Agent Platform. These are not hypothetical commitments. They are live code.
But the spec is v0.9. The registry ecosystem is early. Google’s native support is months away. Companies that invest heavily in ARD catalogs before registries achieve scale are betting on adoption curves they cannot control.
The rational approach is the one the SEO community has converged on for llms.txt: implement what is cheap, monitor what is uncertain, and prioritize what generates measurable results. For ARD, that means tracking adoption through Q3 2026 before committing engineering resources.
Why Searchless Tracks This
At Searchless, we track every standard that affects how AI systems discover, cite, and recommend brands. ARD is early, but the trajectory matters. The web is splitting into two layers: a human-readable layer (HTML, CSS, JavaScript) and a machine-readable layer (structured data, catalogs, knowledge formats). Brands that only optimize for the first layer will be invisible to the agents that increasingly mediate discovery, evaluation, and purchase.
Our AI visibility audit covers the machine-readable layer today. As ARD adoption grows, we will add catalog detection and registry indexing to our tracking. The goal is simple: if an AI system can find your brand, you should know. If it cannot, you should know that too.
FAQ
What is Agentic Resource Discovery (ARD)?
ARD is an open specification published by Google, Microsoft, GitHub, Hugging Face, and seven other companies in June 2026. It defines how AI agents discover and verify tools, APIs, MCP servers, and other agents across the web using publisher-hosted catalog files and crawled registries.
Does ARD replace llms.txt?
No. ARD and llms.txt serve different purposes. llms.txt provides content context for AI crawlers reading your site. ARD defines how agents discover callable capabilities, tools, and services. They address different layers of machine-readable web infrastructure and may coexist.
Should content websites implement ARD today?
Not yet. ARD v0.9 is a draft spec primarily for companies that publish APIs, MCP servers, or AI agents. Content-focused websites have no clear action to take today. The priority remains structured data, answer-first content, and entity authority building.
How does ARD verification work?
ARD uses domain ownership for basic trust. Publishers host an ai-catalog.json file at a well-known path on their domain. For production use, publishers can attach cryptographic trust metadata so agents and registries can verify identity before connecting.
When will ARD impact brand visibility in AI search?
The timeline depends on registry adoption. GitHub shipped an agent finder for Copilot on day one, and Hugging Face released a discovery tool. Google plans native ARD support in its Gemini Enterprise Agent Platform in the coming months. Brands offering tools or APIs should monitor adoption through Q3 2026.
Who created the Agentic Resource Discovery specification?
Eleven companies contributed: Google, Microsoft, GitHub, Hugging Face, Cisco, Databricks, GoDaddy, NVIDIA, Salesforce, ServiceNow, and Snowflake. The spec is licensed under Apache 2.0 and builds on the AI Catalog data model maintained under the Linux Foundation.
The agentic web is not coming. It is here. GitHub Copilot already discovers tools through ARD. Hugging Face already indexes capabilities. The question for brands is the same one we have been asking since search began: can the system find you?
For most companies, the answer is still no. Find out in 60 seconds. Get your free AI visibility score at audit.searchless.ai.