# Method

## Dates

llms.txt: crawl of 2026-09-04, shared with the llms.txt adoption report.
API and MCP probes and the MCP registry download: 2026-09-23.

## Base

6,133 YC domains on the list. 5,630 answered the robots.txt request and count
as reachable. 5,405 of those carry a batch year from 2012 to 2026.

## Sample

150 domains drawn at random from each of the six batch groups (900 in all).
Archived homepages from the 2026-09-04 crawl exist for 892 of them. Every
subdomain and path request went through a rotating proxy; hostnames that did
not resolve were recorded as absent.

## Public API rule

Any of: a homepage link to /api, /api-docs, /api-reference, /docs/api,
/developer(s) or an api., developers. or docs. host with API in the path or
link text; a nav link whose text is "API", "API reference", "API docs" or
"Developers"; a docs., developers., developer. or api. subdomain, or /docs,
that serves a page on its own host containing "API" and two or more of: API
reference, endpoint, API key, curl, OpenAPI, Swagger, bearer, REST API, SDK,
or links to an API reference; an llms.txt that points to API reference pages.

## MCP server rule

Any of: an mcp. subdomain that answers with JSON naming MCP or JSON-RPC, an
HTML page naming its MCP server, or a 400/401/405/406 JSON response; /mcp
answering with JSON-RPC or an MCP page distinct from the homepage;
/.well-known/mcp or /.well-known/mcp.json returning JSON that describes a
server; homepage, docs or llms.txt text naming "MCP server", "Model Context
Protocol", "remote MCP", "via MCP" and similar, or a nav link with MCP in its
text; a listing in the official MCP registry
(registry.modelcontextprotocol.io). Text matches are skipped when negated,
when they describe connecting to other people's servers, or when they are
headlines about MCP security. Wix and Shopify llms.txt template lines about
MCP are counted in `mcp_platform_only`, not as an MCP server (17 sampled
domains).

## Hand checks

20 random positives and 20 random negatives per flag. MCP: first draw one
false positive and one miss, second draw three false positives, third draw
none after fixes. API: first draw one false positive, second draw none.

## Limits

Every flag is a floor. API docs hosted off the company domain and docs sites
that render only in JavaScript are missed. API and MCP shares are from
samples of 150, about 6 points either way per cell.