Data study · Machine-readable doors for AI agents
Only 0.3% of Top Websites Publish an AI Discovery Catalog, and 71.4% of ai-catalog.json Files List an MCP Server
2,202 top-1M homepages publish /.well-known/ai-catalog.json or /.well-known/ard.json, two new files that list a site's AI tools and agents. 93.5% of ai-catalog sites also serve llms.txt; only 9.4% of ard.json sites do.
Analysis of 772,412 top-1M homepages · HTTP Archive, October 2026 crawl
Key findings
- 1
1,058 top-1M homepages (0.14%) publish an ai-catalog.json file and 1,249 (0.16%) publish an ard.json file; 2,202 publish at least one.
- 2
755 of the 1,058 ai-catalog files (71.4%) list an MCP server card, a description of tools an AI assistant can call.
- 3
1,117 of the 1,249 ard.json files (89.4%) list a single entry type, application/ai-registry+json, which points to a shared template.
- 4
93.5% of ai-catalog sites also serve llms.txt, compared with 9.4% of ard.json sites and 14.3% of all homepages.
- 5
Only one top-1,000 site publishes either file; 1,342 of the ard.json and ai-catalog files sit in the 500K to 1M bucket.
Summary
Picture a SaaS company that builds an MCP server so customers' AI assistants can pull their account data. The next question is how an assistant finds out the server exists. One answer is a catalog file at a known address on the company's website.
Two such files now appear in HTTP Archive's crawl: /.well-known/ai-catalog.json and /.well-known/ard.json. Both list AI-facing resources a site offers, such as MCP servers, agent cards, API descriptions or skill files. Neither is a finished web standard yet.
Adoption is tiny but real. 2,202 of the top 1 million homepages publish at least one of the two files, 0.29%. The two groups look different: ai-catalog sites are mostly AI-minded builders who also run llms.txt, while ard.json files look like a template rolled out across many sites at once.
What we measured
We used the HTTP Archive October 2026 mobile crawl (httparchive.crawl.pages) for top-1M homepages. HTTP Archive requests both paths and records whether a file was found, its entry count, the entry types and, for ai-catalog.json, the spec version and host details. Many servers answer any address with a 200 code, so we counted a file only when it had entries, a spec version or host details.
We joined these results with llms.txt adoption and CDN detection from the same crawl.
Two small groups with different profiles
How each file's publishers compare.
| File | Sites | % of top-1M homepages | Also serve llms.txt | % with llms.txt | Served through Cloudflare | % Cloudflare |
|---|---|---|---|---|---|---|
| ai-catalog.json | 1,058 | 0.14 | 989 | 93.5 | 855 | 80.8 |
| ard.json | 1,249 | 0.16 | 117 | 9.4 | 1,206 | 96.6 |
Source: aisearchoptimization.net analysis of 772,412 top-1M homepages · HTTP Archive, October 2026 crawl
For comparison, 57.7% of all top-1M homepages are served through Cloudflare. The near-identical ard.json files, mostly without llms.txt, suggest a provider-level rollout rather than owners writing the file by hand.
MCP servers dominate the catalogs
Entry types found in ai-catalog.json files. A file can list several types, and grouped rows add up the variants, so a site can count more than once in a group.
Sites
| Entry type | Sites listing it |
|---|---|
| MCP server card | 755 |
| Markdown document | 118 |
| Generic JSON | 108 |
| Plain text | 90 |
| OpenAPI description (any variant) | 89 |
| Linkset | 43 |
| Agent skills (any variant) | 94 |
| A2A agent card | 24 |
Source: aisearchoptimization.net analysis of 772,412 top-1M homepages · HTTP Archive, October 2026 crawl
Catalogs mainly announce MCP servers, then plain documents and API descriptions. Only 24 list an A2A agent card, the format AI agents use to describe themselves to each other.
The largest sites have not started
Counts by popularity rank.
| Rank bucket | Homepages | ai-catalog.json | ard.json |
|---|---|---|---|
| Top 1K | 738 | 1 | 0 |
| 1K-5K | 2,920 | 4 | 3 |
| 5K-10K | 3,586 | 5 | 5 |
| 10K-50K | 28,831 | 31 | 23 |
| 50K-100K | 36,937 | 43 | 31 |
| 100K-500K | 306,048 | 366 | 453 |
| 500K-1M | 393,352 | 608 | 734 |
Source: aisearchoptimization.net analysis of 772,412 top-1M homepages · HTTP Archive, October 2026 crawl
Most files sit in the bottom two buckets. These formats are being tried mostly on smaller sites so far, not by large brands.
What this means
Most small businesses do not need these files yet. They matter if you offer something an AI assistant can use directly, such as an MCP server, a public API or a booking tool. In that case, a catalog at a known address is a cheap way to make it findable.
For everyone else, the takeaway is the same as with llms.txt: AI-facing files only help when the basics are in place. A free AI visibility audit from our team tests those basics first, so you know whether a GEO agency should even be talking to you about AI catalogs yet.
Limits
Both formats are early proposals and may change.
HTTP Archive records counts and types, not full file contents.
Only homepage origins are checked.
CDN detection shows where a site is served from, not who created the file.
Embed this research
Add the charts from this study to your page. The embed links back here.
<iframe src="https://aisearchoptimization.net/embed/data-studies/ai-catalog-ard-discovery-files" title="0.3% of Top Sites Publish AI Catalog or ARD Files" width="100%" height="720" style="border:0" loading="lazy"></iframe>
Cite this study
APA
Grenyo, K. (2026). 0.3% of Top Sites Publish AI Catalog or ARD Files. AI Search Optimization. https://aisearchoptimization.net/resources/data-studies/ai-catalog-ard-discovery-files
Plain link
0.3% of Top Sites Publish AI Catalog or ARD Files - AI Search Optimization - https://aisearchoptimization.net/resources/data-studies/ai-catalog-ard-discovery-files
Frequently asked questions
What is ai-catalog.json?
A file at /.well-known/ai-catalog.json that lists AI-facing resources a site offers, such as MCP servers or API descriptions. 1,058 top-1M homepages publish one.
What is ard.json?
A similar discovery file at /.well-known/ard.json. 1,249 top-1M homepages publish one, and 89.4% of them list the same single entry type.
Do I need an AI catalog file?
Only if you offer tools AI assistants can call, such as an MCP server or an API. For most business sites, crawlable pages and clear content matter more.
How common are MCP server listings?
755 of 1,058 ai-catalog.json files (71.4%) list an MCP server card.
Related studies
See where AI search leaves you out
Run the free AI visibility audit. If you want the fixes done for you, the plan is a flat $299/month.