UCP Checker

How we check a store

A check asks a store for its UCP profile the way an agent would, reads the answer against the store's own release, and gives it a status and a score. Here is every step, what each answer leads to, and how the score is made.

What we send

Two requests, both from UCPCheckerBot. Nothing is sent to the store but these, plus the public pages we read for the score.

The profile
GET /.well-known/ucp HTTP/1.1
Host: yourstore.com
Accept: application/json
User-Agent: UCPCheckerBot/1.0 (+https://ucpchecker.com/methodology)
Signature-Agent: sig1="https://ucpchecker.com/.well-known/http-message-signatures-directory"
Signature-Input: sig1=("@authority" "signature-agent";key="sig1");…;tag="web-bot-auth"
Signature: sig1=:…:
How it behaves
  • HTTPS onlyand only to public addresses; a redirect to http or a private network is refused
  • 2 secondsfor the profile, about the time an agent waits
  • Up to 3 redirectseach hop signed for its own host, and every hop shown in the check trace
  • Signed (RFC 9421)Web Bot Auth with an Ed25519 key, so a store can verify the request is ours. About UCPCheckerBot
  • robots.txt firstasked as text/plain; if it disallows us, we don't fetch the profile

The path a check takes

Four steps. A store's answer either passes a step or stops there with a status.

  1. 1

    Read the store's robots.txt robots.txt

    We ask whether UCPCheckerBot may fetch /.well-known/ucp and /. A missing, empty or unreadable robots.txt means yes.

    • Disallowed for UCPCheckerBot Blocked Blocked by robots.txt. We stop and don't fetch the profile.
  2. 2

    Ask for the profile GET

    GET https://{domain}/.well-known/ucp over HTTPS, signed, with a 2 second limit. We follow up to 3 redirects, signing each hop for its own host, so we can show where they lead.

    • 404 Not Detected No profile at the address.
    • 401 or 403 Blocked A firewall or bot wall turned us away. We keep the answer as evidence.
    • 429 No change Rate limited: we learned nothing, so the store keeps its last status (pending if it had none).
    • 5xx, a timeout, DNS or TLS failure, a redirect to http or to a private address Unreachable The store didn't answer in a way an agent could use.
  3. 3

    Read the answer JSON

    A 200 must be a JSON object. The spec says the profile must not redirect. When yourstore.com only forwards to www.yourstore.com (or the other way round) and the profile is served there directly, we judge it there, and the redirect is a To fix. Any other redirect is a spec break. Neither changes the status.

    • Not JSON Invalid The body isn't JSON (UCP_RESPONSE_NOT_JSON).
  4. 4

    Check it against its release Rules

    First the UCP project's own validator (ucp-schema) checks the profile against the schema of the release it declares. That decides whether it is valid. Our rules come on top and can only be stricter.

    • Rejected by UCP's validator Invalid Doesn't match its release's schema. Each error is listed with its path.
    • Missing ucp, version, services or payment_handlers, or a version that isn't YYYY-MM-DD Invalid Present but broken. Each problem is listed with its code and how to fix it.
    • Accepted by UCP's validator, and everything we require is there Verified Verified. Warnings, like missing keys, are listed but don't change it.

A store we have never managed to observe is Pending, not Unreachable: we don't give a verdict we haven't observed. An attempt that observes nothing, like a rate limit, never changes a store's status, score or timing; those always describe the last answer we actually got, and every page says when that was.

What a profile must have

On top of the UCP project's validator, which checks the whole profile against its release's schema. The same at every release since 2026-01-23. (The first release, 2026-01-11, asked for capabilities in place of payment_handlers; we check the few stores still on it against the later rules.) The UCP profile guide explains each part; the UCP Validator runs these rules on a profile you paste.

Required

missing means Invalid
  • ucp.version a release date, YYYY-MM-DD
  • ucp.services at least one, each with its transport and endpoint
  • ucp.payment_handlers present (empty is a warning, not a failure)

Read, but optional

missing means a warning
  • ucp.capabilities what an agent may do
  • keys a JWK Set beside ucp (signing_keys before 2026-08-25); we accept either at every release

Releases we know: 2026-08-25 (current) 2026-04-08 2026-01-23 2026-01-11 . A profile on a release we don't know is checked against the current rules, with a warning.

How we read the web

The market read behind every figure about the wider web on /stats, the home page and the platform pages. It is separate from the daily checks of the stores we check, described above. Quote any figure with credit to UCP Checker and a link to the page and month; anything more needs our permission. Citing and reuse

What we read
Every month we read about 4 million ecommerce sites with measurable shopper traffic (November 2026), the same list each month, not only the stores we check. We ask each one for its UCP profile the way an agent would.
Live sites, and what is left out
About 3.7 million of them were live ecommerce sites. A site we couldn't reach (DNS, certificate, server error or timeout), one that turned automated visitors away, and one that isn't a live store (closed, behind a password, forwarding elsewhere, or not a store) are left out of the live count. Left out, as shares of every ecommerce site read (about 4 million): unreachable 2.5%; turns automated visitors away 4.1%; not live 1%.
Where we look
Every UCP profile lives at one address: /.well-known/ucp on the site's own domain. The path is registered with IANA in the well-known URI registry, the same registry as /.well-known/security.txt, so it is the standard public address any agent asks for. A site that blocks the path, or puts it behind a login, hides its profile from every agent, and from this read; the spec also says an agent must not follow a redirect there, so the profile has to answer on the domain itself.
What counts as a UCP profile
A UCP profile counts when the site serves it publicly at /.well-known/ucp and it passes the official validator; a profile behind a login or a bot wall can't be counted.
How often
Once a month, over the same list of sites, so one month compares with the next. Every figure from the read carries its month; a month-on-month change inside the margin says so.
What each share is of
Every share of the web is of the live ecommerce sites (about 3.7 million live ecommerce sites). Shares of what a profile declares (capabilities, releases, transports) are of the sites with a valid public UCP profile (about 1.2 million sites with a valid public UCP profile). The one chart of every site read is of all sites read, and says so. A true zero reads "none"; a share above 0 and below 0.1% reads "under 0.1%", and one above 99.9% but below 100% reads "over 99.9%", except Shopify, whose shares above 99.9% read 100%: it serves the profile on every store, and the few read without one are edge cases. A group of fewer than 20 gets no share. The stores we check are described by shares only, never counted. We measure what a profile declares, not what a store can do.
Which shares are least certain
We read as a verified, signed bot. The biggest sites are the hardest to reach, so more of them are left out of the live count, and the shares for the top traffic bands are the least certain.

Why the address matters for a store: /.well-known/ucp is now in IANA's registry › · the fields and common mistakes: the UCP profile guide ›

The UCP Score

A number from 0 to 100: three areas, each scored 0 to 100, weighted and added. It describes the last answer we got. Model 1.0.0.

Discovery

Can a UCP agent find and read the profile: HTTPS, an answer at /.well-known/ucp, robots.txt allowing it, and five files outside UCP (llms.txt, sitemap, Open Graph, Organization schema, viewport).

Conformance

Does the profile follow the spec: valid against its release (counted three times), a dated release, keys published, few warnings, spec and schema links that resolve.

Capabilities

What an agent can do here: how many capabilities, checkout, at least one service, payment handlers, and services that answer our check.

Four stores, scored

Scored here by the model itself, before the services check.

Valid profile at the current release with keys, six capabilities including checkout, three transports, a payment handler, and all five files.

100AReady for UCP agents
Discovery × 30%100 → 30
Conformance × 40%100 → 40
Capabilities × 30%100 → 30
A85–100Ready for UCP agents
B70–84Nearly there
C50–69Partly ready
D30–49Hard for UCP agents
F0–29Not ready

A store that's blocked, unreachable or has no profile scores 0 for Conformance and Capabilities. A pending store has no score yet.

What else a check reads

Five files outside UCP

scored, in Discovery
  • /llms.txt answers, isn't empty and isn't an HTML page
  • /sitemap.xml is a real sitemap
  • Open Graph og:title, with og:type or og:url
  • Organization schema JSON-LD naming the business
  • viewport the mobile viewport tag

A file behind a bot challenge page or a rate limit, or a request that got no answer at all, is not assessed and left out of the score, never counted as missing. We also note newer agent files, like agents.md or an MCP server card, when they're there. They aren't scored.

The services check

on a check someone asks for
  • MCP a tools/list call to each declared MCP endpoint
  • REST a read-only search_catalog (or search_locations) call, when the profile declares catalog (or location) search; otherwise the service is shown as declared, not called
  • Links every spec and schema URL in the profile

Read-only: nothing is created, carted or bought. The results count for 20% of Conformance (links) and of Capabilities (services).

One store, one address

Whatever is typed, a store has one page. We trim it, drop the scheme, the path and a leading www., drop a pasted profile filename, and lowercase it.

https://www.YourStore.com/products/hatyourstore.com/status/yourstore.com

What it costs

The check, the score, the status pages and the validators are free, with no sign-up. To be gentle on stores, each store is checked live at most once every 10 minutes, whoever asks, our own daily check included: inside that window you see the latest result and when it can be checked again. A shared limit on live checks across the site keeps them fair for everyone; when it's reached, we ask you to try again in a minute.

Questions

Do you check my store without asking?
Unless you give us access, we read only what your store publishes: the UCP profile, robots.txt and public pages, the same files any agent reads. Pre-production endpoints are checked only when you give them to us. Every request names UCPCheckerBot and is signed, so you can verify it came from us. If robots.txt disallows us, we stop.
Why is my store "Unreachable" when it loads in my browser?
Browsers wait longer and look different to a firewall. We give the profile 2 seconds, the time agents usually allow, and treat 5xx answers, timeouts and DNS or TLS failures as unreachable. Check the store again to see the exact answer we got.
Does a redirect fail the check?
It doesn't change the status: we follow up to 3 redirects and judge what's at the end. The spec says the profile must not redirect. If your bare domain only forwards to www. (or the other way round), we work out where your store lives from where your homepage lands, then check its canonical address and where your profile's services point, and judge the profile there; the redirect becomes a To fix, and the finding lists that evidence. Any other redirect is a spec break. If a redirect hides a profile that is really there (for example, a rule that adds a trailing slash), we say so.
Is response time part of the score?
No. We record how long the profile took and show it, but the score doesn't use it.
Why did my score change when nothing changed?
Usually the services check: when someone asks for a fresh check, we also call each declared service and fetch each spec and schema link, and those results count for 20% of Conformance and Capabilities. A link that was down then, or a service that answered differently, moves the score.
Can I see the raw evidence?
Yes. Each status page has a check trace: every request we made, with its status, size and timing, grouped by the way an agent reaches the store. It also shows the profile as it was and every problem with its code. The UCP Validator runs the same rules on a profile you paste.
What if a bot challenge answers instead of my page?
Some sites answer automated requests with a challenge page and a 200. We recognise the common ones. A file or page behind one is "not assessed", not missing, and it doesn't count against the score, just as when we're rate limited.

Explore protocol data