Dmytro Galko · open data

What AI shopping agents actually see

A census of 5,356 live Shopify storefronts · 2026-09-18

Google and Shopify published the Universal Commerce Protocol so an agent can read a shop, find a product and buy it. The obvious question is how many shops actually speak it — and nobody had published a count. Here is one, with the tool, the domain list and the raw results, so anyone can repeat it and get their own number.

The short answer surprised me: adoption is not the problem. 98.9% of these shops publish a valid profile, all of them on the current 2026-08-25 release, none with a blocker. Shopify turned it on for everyone, so "does my shop support it" is already answered. The real variation is one capability — and it is the one that decides whether the agent is a stranger or your customer.

Half of them treat the agent as a stranger

Agent can sign the shopper into their account2,78452.5%
Agent can only act as a guest2,51547.5%

dev.ucp.common.identity_linking is the only capability these 5,299 shops differ on; the other eight are identical everywhere. Without it the agent can browse and buy, but as an anonymous visitor: no saved address, no order history, no loyalty tier, no "the usual". For a returning customer that is the difference between a checkout and a form.

What 5,356 shops answered

Profile read and audited5,29998.9%
No profile published at all110.2%
Refused the request (403, 429)130.2%
Answered something that is not a profile10.0%
No answer320.6%

46 shops (0.9%) could not be checked either way. They are counted here and nowhere else: a shop that refused the request is not a shop that failed the audit, and none of them appears in any percentage above or below.

What an agent is allowed to do

How often each capability appears among the 5,299 audited shops.

dev.shopify.catalog
use Shopify's own catalogue extension
5,299100.0%
dev.ucp.shopping.cart
put it in a cart
5,299100.0%
dev.ucp.shopping.catalog.lookup
look a product up by id
5,299100.0%
dev.ucp.shopping.catalog.search
find a product by searching
5,299100.0%
dev.ucp.shopping.checkout
complete the purchase
5,299100.0%
dev.ucp.shopping.discount
apply a discount code
5,299100.0%
dev.ucp.shopping.fulfillment
see delivery options
5,299100.0%
dev.ucp.shopping.order
read the order afterwards
5,299100.0%
dev.ucp.common.identity_linking
sign the shopper into their own account
2,78452.5%

Eight of the nine are dev.ucp.*, the protocol's own namespace. The ninth, dev.shopify.catalog, is a platform extension — evidence that the extension mechanism is being used in production and not only in the specification.

Which version they publish

2026-08-255,299100.0%

Not one audited shop is behind the 2026-08-25 release. A platform-wide rollout has no long tail of old versions, which is the opposite of what a self-hosted protocol looks like a year in.

Blockers found: none

The audit looks for twelve ways a profile can be published and still be useless to an agent — a checkout with no searchable catalogue, a capability declared without the schema needed to call it, an extension naming a parent that is not there, an endpoint on plain HTTP. Across 5,299 shops it found none of them. Worth stating plainly: the tool is not measuring nothing — the same twelve checks do fire on hand-written profiles, which is what they were written against.

How it was measured

What this does not say

Want this run against your own shops, or the identity-linking gap closed on yours? Write to hello@dkautomation.dev or open an issue.