Known Good census — September 2026

Two website platforms switched on a live agent endpoint for every site they host, and the standard MCP check finds none of them. That is the September finding from Known Good, the index of websites whose agent-facing capabilities are probe-verified rather than declared. Everything below was measured on a stated date over a stated population, and is reproducible from the method at the end.

Which website platforms give AI agents an endpoint?

Wix and Shopify do, for every tenant. Eight others do not. On 17 September 2026 we probed 50 randomly sampled tenants on each of ten platforms, custom domains only, one date, one probe. Every Wix tenant and every Shopify tenant that could be asked answers tools/list on a live MCP endpoint.

An agent-actionable endpoint is a strict MCP tools/list that answers, or a published API catalog. It is a probe result, not a reading of the site's copy. Shopify's four missing tenants are three whose robots.txt refused and one that could not be reached; they sit beside the denominator, not inside it.

Platform-supplied means one decision, not fifty. Wix's endpoint is on wixapis.com; Shopify's on {shop}.myshopify.com. Each platform registers one identical tool signature for every tenant, and Wix's eight tool names are a strict subset of Shopify's thirteen, in the same order. That is consistent with one published list, not two designs. BigCommerce is the mirror image: 45 of 50 serve markdown when an agent asks for it, and none is callable. Readable is not the same as callable.

Why the MCP server card misses most live endpoints

None of the 96 was found through the MCP server card. All 96 were found through /.well-known/ucp on the tenant's own domain, the Universal Commerce Protocol profile, which advertises the MCP endpoint the platform runs. A check that reads only the card records every one of these sites as having no MCP server.

That included Known Good's own check until 4 September 2026. Re-probing 1,284 ecommerce sites already in the index through the UCP route found 23 live endpoints the card route had recorded as absent: 20 on Shopify's infrastructure, three on the merchant's own host. The index now records every route it tried, so adding a route reads as a method change rather than a silent count movement.

What is Agentic Resource Discovery (ARD)?

ARD is a discovery standard that lets a website tell AI agents what resources it exposes and what questions each one answers. It was published on 26 August 2026 as v0.91, a proposal, authored jointly at Google, Microsoft and Hugging Face.

How a site publishes ARD. Three mechanisms, any of which an agent can find without a prior link:

• /.well-known/ard.json, a JSON file with an entries array
• Agentmap: https://…/entries.json, a directive in robots.txt
• <link rel="ard"> in the page head, with rel="ai-catalog" kept as a predecessor for compatibility

Each entry carries an identifier, a display name, a media type, a URL or inline data, and representativeQueries: two to five natural-language examples of what the resource answers. That field is a site telling an agent what to ask it, in the site's own words.

How many sites use ARD? No published count exists yet; the standard is four weeks old. Known Good's October census will carry the first, measured from stored robots.txt directives and page headers across the index and dated to the standard's first weeks.

One implementation note. At least one widely used readiness scanner checks the predecessor rel="ai-catalog" and not the current rel="ard". Implementing only what a scanner greps for is the wrong lesson. Serve all three mechanisms and both link relations; knowngood.sh does, at /.well-known/ard.json and /.well-known/ai-catalog.json.

How many sites use WebMCP?

580 sites register at least one WebMCP tool when the page runs, out of 14,351 in the index. WebMCP is the browser-side standard that lets a page register tools an agent can call directly, through navigator.modelContext or document.modelContext. We measured it in two layers on 30–31 August 2026: a static pass for the code, then a runtime pass in a real browser for what actually registers.

Most WebMCP adoption is one plugin. Treating each site's exact set of tool names as a fingerprint puts 39.5% of the 580 in a cluster with at least one other site; the largest cluster, 18.1%, is one WordPress plugin's default. The most-registered tools are search_site (122 sites), fetch_as_markdown (105), search_content (28), get_contact_info (18), get_pricing and search_products (17 each).

Four things a WebMCP count has to get right.

1. Match a call, not a mention. An article explaining WebMCP is not a site using it.
2. Match both namespaces. The spec renamed navigator.modelContext to document.modelContext during its origin trial. A detector reading one namespace undercounts.
3. Run the page. 647 sites carry the code; 580 register something. The gap is code that ships and never executes.
4. Count "could not ask" separately. 1,224 sites' robots.txt disallowed a script we needed. Every other WebMCP checker we found reports those as failures. They are not failures; they are unknown.

There are more WebMCP checkers on the web than sites using WebMCP. Known Good's, at knowngood.sh/verify/webmcp, is the one that reports a site's position in a measured population, and it declares its own tools so an agent calls it rather than reads it. It currently detects the imperative API and not the declarative toolname attribute form; that count follows.

The standing figures

Rubric v1.1, frozen 18 September 2026. Each figure carries the date it was last re-graded and the population it was measured over.

Of 6,026 MCP servers that answer, 5,371 are two platforms' shared tooling. 632 were built by the site itself. Of 10,877 sites serving markdown, 4,372 are one documentation platform's default. A platform moving is not thousands of site owners deciding, and every figure's page says which it is.

Who reads an index built for agents?

AI crawlers outnumber Googlebot on this index by at least nine to one on every day measured, and the reader rotates. In its first fortnight the index was swept in turn by each major AI corpus crawler, Anthropic's, then Amazon's, then Perplexity's, each taking the majority of requests for a few days and leaving. A sweep is ingestion. It is flattering and inert.

The path that matters is a person asking. Claude-User and ChatGPT-User requests mean someone wanted an answer. Two-thirds of Claude-User fetches take the markdown twin rather than the HTML page, and those twin URLs appear nowhere but our own Link: rel="alternate" header. Agents find the alternate and follow it. Agents read this index through other people's search engines; they do not yet query it directly, and this census says so rather than inflating a request count into usage.

Method

• Consent before every fetch. robots.txt first. ai-input=no, a Disallow for our agent, or noai stops the probe; the site is counted as consent-excluded, beside the denominator. A robots.txt that errors is treated as disallow.
• We identify ourselves. Every request carries the user-agent KnownGood-Verifier and a link to the opt-out page. Retry-After is honoured.
• We list tools; we never invoke them. initialize and tools/list, nothing further. Invoking a stranger's tool could place an order.
• Unknown is not no. Every count has three states: yes, no, and could not ask.
• A sample is a sample. 50 per platform, one date, one vantage, a 95% Wilson interval beside every rate. Tenant lists come from a public technology-detection dataset, August 2026 snapshot.
• Every figure carries its date. A figure that moves is corrected as a dated row in the public changelog, never edited in place.

Frequently asked

Does Shopify give AI agents an MCP endpoint? Yes, for every store, as of 17 September 2026. It is served from the store's myshopify.com host and advertised through the store's /.well-known/ucp profile, not through an MCP server card.

Does Wix? Yes, for every site, same date, served from wixapis.com through the same route.

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.

Can an agent complete a purchase on these sites? This census does not say so. An endpoint that answers tools/list is callable; completing a checkout needs checkout tools in the signature and, strictly, a completed action. That is a separate measurement.

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.

Next

The October census runs on the same rubric, frozen at v1.1, so its figures are a series point against August's and September's. It carries the first measured counts for ARD, the A2A agent card and the agent-skills index, and the re-probe of listed Wix and Shopify sites.

The data

Everything above is in the public census repository, and this page is the canonical version of it: census/2026-09.md is the source document.

All censuses · How we grade · Browse the index