# Known Good census — October 2026

Home > Census > October 2026

By Dean, Known Good · Published 8 Oct 2026 · Last updated 6 Nov 2026 · Figures as of 2 Oct 2026

**57,266 websites can be read or acted on by an AI agent, out of 95,529 we asked.** That is the October finding from Known Good, the index of websites whose agent-facing capabilities are probe-verified rather than declared. It is the first census over the whole index rather than a sample: one pass, one rubric, every host asked the same way between 26 Sep 2026 and 28 Sep 2026, with a markdown re-check on 2 Oct 2026. Everything below was measured on a stated date over a stated population and is reproducible from the method at the end.

## How many websites are agent-ready?

**57,266.** We started from 104,578 hosts in scope, asked 95,529, and found something built for agents answering on 83,917. 57,266 met the rubric. 606 hosts declined and 2,382 could not be asked; both stand beside the denominator, never inside it.

| Stage | Hosts |
| --- | --- |
| In scope | 104,578 |
| Asked | 95,529 |
| Declined (robots.txt or an AI-use signal) | 606 |
| Could not be asked | 2,382 |
| A valid door answered | 83,917 |
| Listed under rubric v1.2 | 57,266 |

A site is listed when at least two of its agent-facing signals are present and at least one of them is strong. The strong signals are markdown served on request, markdown at a markdown URL, an MCP endpoint that returns its tool list, and six files at `/.well-known/` paths: `api-catalog`, `mcp/server-card.json`, `agent-skills/index.json` (the Agent Skills index), `oauth-authorization-server` and `oauth-protected-resource` (the OAuth pair, counted only beside an answering MCP or A2A endpoint), and `http-message-signatures-directory` (the message-signatures directory). The medium signals are `llms.txt`, AI agents named in robots.txt and Link headers; each counts as a second signal and never lists a site by itself. The rubric is published at [/benchmark](/benchmark); the probe's behaviour at [/bot](/bot).

## How many websites answer an agent's tool calls?

**48,819 answer a tool-list request from any agent, with no sign-in.** That is the strict count: an MCP server that answers `initialize` and `tools/list` to a caller it has never seen. 790 server cards point at nothing and are not counted among answering endpoints; a card is a claim, an answer is a measurement.

## Who built those servers?

**The platforms did, for 46,427 of the 48,819.** When a website platform switches on an agent endpoint, every site it hosts becomes callable on the same day, and that is where most of the web's agent-readiness has come from so far. A server counts as platform-supplied when the same tool set is served to at least 200 unrelated domains, when it sits off the site's own domain, or when it carries one commerce platform's storefront fingerprint. The other 2,392 were built by the site owners themselves, one decision each, and those are the servers listed at [/mcp-servers](/mcp-servers).

| Source of the answering server | Sites |
| --- | --- |
| Platform-supplied, all platforms | 46,427 |
| One documentation platform's search | 4,484 |
| One course platform's catalogue | 309 |
| Built by the site | 2,392 |

September's sample found Wix and Shopify had done this for every tenant. The full pass puts a number on what that means across the web: most sites that answer tool calls today do so because their platform decided to, and the index labels each one so a reader can tell a platform's decision from a site owner's.

## What changed in the rubric?

**Platform-supplied is now a measured class with three criteria, and a discovery file is no longer one of them.** Rubric v1.2 defines platform-supplied by the shared tool shape (200 unrelated domains), by an endpoint off the site's domain, or by the storefront fingerprint. Under v1.1, finding a server through a `/.well-known/ucp` profile had been treated as a platform signal; it is a route in, not evidence of who built the server, and 49 hosts moved from platform-supplied to self-built when the criterion was withdrawn. The change is dated on [/benchmark](/benchmark).

## How many sites use WebMCP?

**Up to 992 sites carry a WebMCP call in the page; the runtime count over the full index follows in November.** The static count is an upper bound: the probe stores the first 8 KiB of a page, so a call seen beside the namespace can be counted but not re-judged without a fetch. 5,956 sites did not return their page to the probe and are counted as unknown, never as failures. The last runtime pass, on 31 Aug 2026 over the index as it then stood, recorded 580 sites registering through WebMCP; 579 of them registered at least one tool, and one registered an empty list.

## What is x402, and how many sites use it?

**x402 is an open payments protocol built on HTTP: a resource that requires payment answers with status 402 and a `PAYMENT-REQUIRED` header stating the price and how to pay, and the client retries with a `PAYMENT-SIGNATURE` header.** The full page is [/standard/x402](/standard/x402). In the full pass 218 sites answered 402 to a probe and 61 are listed; 52 publish the x402 schema and 44 are listed. A 402 alone is a door; the schema is the door with a handle on it.

## What is ARD, and how many sites publish it?

**355 sites publish Agentic Resource Discovery by at least one of its mechanisms, out of 92,541 whose robots.txt and page head the probe could read.** ARD lets a site tell an agent what resources it exposes and what each one answers; September's census described the standard at v0.91 and promised the first measured count. This is it, read from stored robots.txt files and page heads on 28 Sep 2026.

| Mechanism | Sites |
| --- | --- |
| [`/.well-known/ard.json` referenced](https://knowngood.sh/standard/ard.md) | 111 |
| [`Agentmap:` directive in robots.txt](https://knowngood.sh/standard/ard.md) | 241 |
| [`<link rel="ard">` in the page head](https://knowngood.sh/standard/ard.md) | 90 |
| [`rel="ai-catalog"` without `rel="ard"`, the predecessor alone](https://knowngood.sh/standard/ard.md) | 172 |
| [Any mechanism](https://knowngood.sh/standard/ard.md) | 355 |

Beside the total: 5,947 sites were read at robots.txt only and showed no ARD signal there, 2,382 could not be asked, and 9,049 were in scope and not asked in this pass. A standard 4 weeks old on the measurement date is published by a few hundred sites; the page at [/standard/ard](/standard/ard) carries the count.

## How do sites get into the index?

**Four routes, one probe, and placement is never a route.** A site declares agents welcome in its robots.txt; a site publishes a file or endpoint meant for agents; a site appears in a public MCP or API registry; or an owner asks at [/submit](/submit). 13,284 domains came from public registries; 2,655 of the listed hosts arrived by that route.

## Who reads an index built for agents?

**AI crawlers out-request Googlebot on every day of the window, by 30 to one on the quietest day and by more than a thousand to one on the sweep days.** Over the eleven days from 26 Sep to 6 Oct, AI crawlers made 1,158 times Googlebot's requests in total. AI crawlers here are the user agents ChatGPT-User, OAI-SearchBot, GPTBot, Claude-User, Claude-SearchBot, ClaudeBot, Perplexity-User, PerplexityBot, Amazonbot, meta-externalagent, meta-externalfetcher, DuckAssistBot, MistralAI-User, YouBot and Bytespider; search engines (Googlebot, bingbot, Applebot) are not among them. One corpus crawler took about 87,900 requests on a single day and about 53,500 on another; another took about 27,900 on its busiest day. A sweep is ingestion. It is flattering and inert.

**The fetches that mean a person asked arrived every day.** `ChatGPT-User` requests, which carry a person's question behind them, came on all eleven days; `Claude-User` on 8 of them. September's finding holds: agents reach this index through other people's search engines and through a person asking, and the difference between a sweep and a question is visible in the log.

The MCP endpoint at `/mcp` received about 1,600 requests a day across the window, steady from day to day. This source counts requests by path and cannot say who made them, so the census states the count and does not call it agent use; the split by caller is a separate read. Counts in this section are provider estimates over all statuses; sampled counts are rounded and said to be approximate.

## Method

- **Consent before every fetch.** robots.txt first; `ai-input=no`, a `Disallow` for our agent, or `noai` stops the probe and the site is counted as declined, beside the denominator. A robots.txt that errors is treated as disallow.
- **The probe is public.** Its name, the paths it reads, its rate, and how to opt out are published at [/bot](/bot).
- **We list tools; we never invoke them.** `initialize` and `tools/list`, nothing further.
- **Unknown is not no.** Every count has three states: yes, no, and could not ask.
- **Every host is re-probed each pass.** A site that has lost a capability since the last pass keeps its record.
- **Every figure carries its date.** A figure that moves is corrected as a dated row in the public changelog, never edited in place.
- **Requests made in this pass:** 2,149,241.

## Next

The November census runs on rubric v1.2 unless [/benchmark](/benchmark) records a change before then. It carries the WebMCP runtime count over the full population, the A2A agent-card count, and the re-probe of the 174 hosts whose markdown check returned no verdict in this pass.

## Frequently asked

**How many websites have an MCP server?** 48,819 of the sites we probed return their tool list without credentials, as of 28 Sep 2026. A further 790 publish a server card that points at nothing; they are not counted among answering endpoints.

**Is a Shopify or Wix store "agent-ready"?** It answers tool calls, because the platform runs a server for every store. Known Good lists it and labels the server platform-supplied, so a reader can tell one platform's decision from a site owner's.

**What is the difference between declared and verified?** Declared means the site says it has a capability: a card, a header, a file. Verified means we asked and it answered. Known Good lists only what answered, and shows the date.

**Why is this census over the whole index and September's over a sample?** September probed 50 tenants on each of ten platforms on one date, under rubric v1.1. October is the first full pass under rubric v1.2, and every host was asked. The two are series points on the same rubric lineage, with the v1.2 change dated above.

**Where do the numbers come from?** From Known Good's own probes, each dated on the page it appears on, with the population it was measured over stated beside it. The data behind every count is reachable at /api/stats and on each site's report page.

## The data

Everything above is in the public census repository, and this page is the
canonical version of it: https://knowngood.sh/census/2026-10

- Source document: https://github.com/knowngood-sh/agent-readiness-census/blob/main/census/2026-10.md
- the funnel: in scope, asked, declined, could not be asked, answered, listed: https://github.com/knowngood-sh/agent-readiness-census/blob/main/census/2026-10-funnel.csv
- where the listed sites' answering MCP servers come from: https://github.com/knowngood-sh/agent-readiness-census/blob/main/census/2026-10-mcp-sources.csv
- Agentic Resource Discovery, by mechanism: https://github.com/knowngood-sh/agent-readiness-census/blob/main/census/2026-10-ard.csv

Prior census: September 2026, https://knowngood.sh/census/2026-09.md
All censuses: https://knowngood.sh/census/