---
title: "The ard.json File: Does Your Site Need It?"
description: "The new /.well-known/ard.json file describes tools for AI agents, not your content. Across the logs of 8 sites, it was requested mostly by Lighthouse."
date: 2026-11-21
author: "Ewelina Podrez-Siama"
url: "https://podrez.pl/en/ard-json/"
categories:
  - "SEO Blog"
tags:
  - "AI search &amp; GEO"
  - "Technical SEO"
---

As of November 21, 2026 · logs from 8 sites, November 1–20, 2026

Guides on “AI agent readiness” have a new file on their to-do lists: `/.well-known/ard.json`. Before that, `ai-catalog.json` played the same role. It sounds like another [llms.txt](https://podrez.pl/en/llms-txt/), so the familiar question is back: does my site need it?

The short answer: if you run a blog, a company website or an online store without its own API and MCP server, you don’t need ard.json today. The file doesn’t describe your content. It’s a list of tools an AI agent can call: MCP servers, agents, skills and registries. A site without tools like these simply has nothing to put in it.

What ARD is

**Name:** Agentic Resource Discovery, a specification for how an AI agent finds the tools a company or website makes available.

**Who:** the authors are Junjie Bu (Google), R.V. Guha (listed as a Technical Fellow at Microsoft in Microsoft’s announcement) and Shaun Smith (Hugging Face). Announced on June 17, 2026, simultaneously on the Google, Microsoft and Hugging Face blogs.

**Status:** v0.91, dated August 26, 2026, marked “Proposal.” So it’s a proposal, not an established standard.

**Where:** the `/.well-known/ard.json` file or a `<link rel="ard">` tag in the page code.

**What it describes:** MCP servers, A2A agent cards, skills and registries. It doesn’t list articles, products or categories.

[The ARD specification](https://agenticresourcediscovery.org/spec/)



**194**requests for ard.json and ai-catalog.json

**192**of them came from Lighthouse and WebPageTest audits

**2**requests from Andi-SearchBot, one for each file

**1,583**traffic-advice requests from Chrome

 

Server logs, November 1–20, 2026, 8 sites I run: an SEO blog, a food blog and travel guides in Polish, English and German. I don’t have any of these files, so almost every response was a 404.



Table of Contents

[Toggle](#)

- [Who requests ard.json? What my logs show](#who-requests-ardjson-what-my-logs-show)
- [ard.json or ai-catalog.json?](#ardjson-or-ai-catalogjson)
- [What’s inside](#whats-inside)
- [By contrast: traffic-advice, a file with a real reader](#by-contrast-traffic-advice-a-file-with-a-real-reader)
- [When ard.json makes sense](#when-ardjson-makes-sense)
- [What to do instead if your site is about content](#what-to-do-instead-if-your-site-is-about-content)
- [Frequently asked questions](#frequently-asked-questions)



Who requests ard.json? What my logs show
----------------------------------------

Across the logs of the 8 sites I run, ard.json and ai-catalog.json were requested 194 times between November 1 and 20. My first thought? Agents are already looking for it! Then I checked: 192 of those requests were Lighthouse and WebPageTest audits. 130 came from Lighthouse running on Google’s servers (that’s how Google’s tools such as PageSpeed Insights work), and 62 from WebPageTest. The Lighthouse requests identify themselves as Chrome-Lighthouse in the logs, not Googlebot, so a Google IP address doesn’t mean the search engine has started reading ard.json.

Who ran those audits? Mostly me, testing my own sites. The rest, I suspect, came from SEO friends 😉 The lesson applies to any log analysis: before you decide that agents are looking for a new file, make sure it isn’t just you running audits.

Why is Lighthouse in this story at all? Starting with version 13.5 (September 2026), it includes an ARD audit in the experimental Agentic Browsing category (testing it requires Chrome 150 or later). Starting with version 13.6 (October 27), it looks for the manifest the way the v0.91 specification describes: in robots.txt, in a `rel="ard"` link, in an HTTP header, at `/.well-known/ard.json`, and finally in the legacy locations (ai-catalog.json and `rel="ai-catalog"`), which trigger a warning. That’s why requests for the two files show up in pairs in the logs. If it finds nothing, the audit is marked “not applicable” and doesn’t affect the score. The same category also checks llms.txt.

Bots were few and far between. On November 3, Andi-SearchBot (from the Andi search engine) requested both files on podrez.pl along with a dozen or so other agent-related URLs. AgentTrustBot, which builds a directory of agents, requested agent cards and MCP server cards on four of my travel guides on November 5 and 6, but not ard.json. The ChatGPT, Claude, Gemini and Perplexity bots didn’t request the file at all.

I also couldn’t find any documentation in which OpenAI, Anthropic, Google or Microsoft say their assistants read ard.json from websites. Google has announced native ARD support for Agent Registry on Gemini Enterprise Agent Platform, a tool for businesses. It’s due in the coming months. The announcement doesn’t mention Google Search. The “Connect a chatbot” guides on the specification site hook a tool search service (GitHub Agent Finder or Hugging Face Discover) up to Claude, ChatGPT or Copilot. According to the specification, a registry crawler is supposed to fetch the file from your site, and the chatbot then queries that registry. That’s how Agent Finder in GitHub Copilot works, for example: it searches for tools in a registry chosen by you or your company. This matters for measurement. My logs show who requested the file on my server, but any later query from an assistant to a registry happens outside my domain.

The specification’s acknowledgments list people from AWS, Cisco, Databricks, GitHub, GoDaddy, Google, Microsoft, Nvidia, Salesforce and Snowflake. The authors thank them for their contributions and feedback, but this isn’t a list of companies that have implemented the file.

ard.json or ai-catalog.json?
----------------------------

If an AI visibility guide tells you about `ai-catalog.json`, that’s not a mistake, just an older version of the file. ARD’s predecessor used the `/.well-known/ai-catalog.json` path. Since v0.91, there’s one path: `/.well-known/ard.json`. Tools may still check the old path (“MAY” in the specification), but publishers should move the file to the new one (“SHOULD”). Some guides still describe the old name.

Besides the file, the specification describes four more discovery methods: a `rel="ard"` link in the `<head>`, an `Agentmap:` line in robots.txt, JSON-LD entries on the page that describes the tool, and DNS records.

What’s inside
-------------

The file is JSON with one required array, `entries`. Each entry describes one tool and must have four things:

- `identifier`, an ID in the `urn:air:domain:name` format,
- `displayName`, a human-readable name,
- `type`, the type of description, for example `application/mcp-server-card+json` for an MCP server card or `application/a2a-agent-card+json` for an agent card,
- `url` or `data`: the address of the description or the description itself, inline (exactly one of the two).

`representativeQueries` are also recommended: 2 to 5 sample queries that should prompt an agent to reach for the tool. Optional fields include a description, tags and a version.

By contrast: traffic-advice, a file with a real reader
------------------------------------------------------

The same logs contain a .well-known file that gets requested every day: `/.well-known/traffic-advice`. Over 20 days, Chrome requested it 1,583 times across all 8 sites. The requester is Chrome’s Private Prefetch Proxy. Before the proxy sets up a connection for a page being prefetched, for example from Google search results, it checks this file to see whether the site wants to limit that traffic.

On my sites the response is a 404, and it’s staying that way, because a missing file means the proxy uses its default behavior. In [traffic-advice in your logs](https://podrez.pl/en/traffic-advice/), I explain what the file does, what else showed up in the logs and when it’s worth adding. But the difference is already clear: one file has a real reader, and the other is read mostly by Lighthouse.

When ard.json makes sense
-------------------------

- **You have a public MCP server** (for example, with product search, bookings or data) and you want agents to find it.
- **You offer an A2A agent** with an agent card, or a skills package.
- **You have an API for agents** and describe it with an MCP server card or an agent card.
- **You run a registry or directory** of tools for agents.

Here’s what the file could look like for a food blog with its own MCP server for recipe search. This is my example, not part of the specification:

```
{
  "entries": [
    {
      "identifier": "urn:air:example.com:mcp:recipes",
      "displayName": "Recipe search",
      "type": "application/mcp-server-card+json",
      "url": "https://example.com/.well-known/mcp/server-card.json",
      "description": "MCP server for searching recipes.",
      "representativeQueries": [
        "gluten-free bread recipe",
        "30-minute keto dinner"
      ]
    }
  ]
}
```

Note the `url`. It points to an MCP server card, which is another file you need to have. And the card points to a working server. Without one, ard.json is a signpost to nowhere.

Even then, keep in mind that this is still a v0.91 proposal. The types aren’t registered with IANA yet, and the authors note that the format may change, so treat any implementation as a test.

Check your site

You can check whether your site has ard.json or the older ai-catalog.json with my tool **Can AI bots read your site?** It shows whether the file exists, how many entries it has and what types they are (for example, an MCP server). It also flags entries missing required fields. The result is informational and doesn’t change the bot access ratings.

**Example from November 21, 2026:** github.com doesn’t have ard.json, but it does have the old ai-catalog.json with one MCP server. That entry is missing the `displayName` field, so by ARD’s rules it’s incomplete.

[Check ard.json on your site](https://podrez.pl/en/robots-txt-checker/)



What to do instead if your site is about content
------------------------------------------------

- **Check whether AI bots can read your site at all.** The [Can AI bots read your site?](https://podrez.pl/en/robots-txt-checker/) tool shows what robots.txt, meta tags and headers allow AI bots and search engines to do.
- **Decide on model training.** I described one robots.txt rule for this in my post on [\_\_usage\_ai-training](https://podrez.pl/en/usage-ai-training/).
- **Look at your logs.** [Server log analysis](https://podrez.pl/en/server-log-analysis-ai-visibility/) starts with who actually shows up. You can also check AI bots in your own logs with the [AI bots in logs](https://podrez.pl/en/ai-bots-in-logs/) tool.

To sum up: ard.json is a business card for tools, not for content. If your site has nothing to hand to agents, the file will stay empty, and its main visitor will be Lighthouse.

Frequently asked questions
--------------------------

### What is the ard.json file?

It’s a manifest from the Agentic Resource Discovery (ARD) specification, published at `/.well-known/ard.json`. It describes tools an AI agent can call: MCP servers, A2A agent cards, skills and registries. It doesn’t describe your site’s content.



### Does ard.json improve visibility in ChatGPT or Google?

There’s no evidence that it does. I couldn’t find documentation showing that ChatGPT, Claude, Gemini or Copilot read this file from websites (GitHub Copilot can search for tools through Agent Finder, but then it queries a registry, not your site). In the logs of my 8 sites from November 1–20, 2026, no ChatGPT, Claude, Gemini or Perplexity bot requested it. Lighthouse checks it in the experimental Agentic Browsing category, but a missing file doesn’t affect the score.



### What’s the difference between ard.json and ai-catalog.json?

ai-catalog.json is the path from ARD’s predecessor. Since v0.91 (August 26, 2026), the path is `/.well-known/ard.json`. Tools may still check the old path, but publishers should switch to the new one.





 **Sources** J. Bu, R.V. Guha, S. Smith, [Agentic Resource Discovery Specification v0.91](https://agenticresourcediscovery.org/spec/), August 26, 2026.
 GitHub ards-project, [ard-entry.schema.json schema](https://github.com/ards-project/ard-spec/blob/main/spec/schemas/ard-entry.schema.json), accessed November 21, 2026.
 Google for Developers Blog, [Announcing the Agentic Resource Discovery specification](https://developers.googleblog.com/announcing-the-agentic-resource-discovery-specification/), June 17, 2026.
 Microsoft (R. Guha), [Introducing the Agentic Resource Discovery specification](https://commandline.microsoft.com/agentic-resource-discovery-specification-ard/), June 17, 2026.
 Hugging Face, [Agentic Resource Discovery: Let agents search for tools, skills, and other agents](https://huggingface.co/blog/agentic-resource-discovery-launch), June 17, 2026.
 GitHub Changelog, [Agent finder for GitHub Copilot now available](https://github.blog/changelog/2026-06-17-agent-finder-for-github-copilot-now-available/), June 17, 2026.
 GoogleChrome/lighthouse, [releases 13.5.0 and 13.6.0](https://github.com/GoogleChrome/lighthouse/releases), September 17, 2026 and October 27, 2026.
 ARD, [Connect a chatbot](https://agenticresourcediscovery.org/connect/), accessed November 21, 2026.
 Chrome for Developers, [ard.json](https://developer.chrome.com/docs/lighthouse/agentic-browsing/ard), updated October 20, 2026.
 Chrome for Developers, [Lighthouse agentic browsing scoring](https://developer.chrome.com/docs/lighthouse/agentic-browsing/), updated May 5, 2026.
 Chrome for Developers (K. Hempenius, K. Baheux, M. Buettner), [Private prefetch proxy in Chrome](https://developer.chrome.com/blog/private-prefetch-proxy), updated May 11, 2022.
 Private Prefetch Proxy, [Traffic Advice](https://github.com/buettner/private-prefetch-proxy/blob/main/traffic-advice.md), proposal explainer, accessed November 21, 2026.
 Own data: server logs from 8 sites, November 1–20, 2026, accessed November 21, 2026. 

An LLM may have supported me in preparing this text – most often at the translation, research, proofreading or code-styling stage. The responsibility for the decisions, the claims made and the arguments cited is fully mine. More on [how I work with AI](https://podrez.pl/en/authorship-and-ai/).