Agent-Readiness of 50 Cross-Border DTC Brands (2026)

What it found

• 25 / 47

  reachable storefronts serving a UCP manifest

  53.2%

• 25 / 25

  of those manifests issued by Shopify, not the brand

  22 live on 2026-08-21 · 3 by host hint

• 0 / 47

  exposing an agent-payments endpoint

  no exceptions

• 1 / 47

  exposing an MCP endpoint

  2.1%

• 9 / 47

  with Product schema on the sampled product page

  19.1%

• 6 / 47

  naming any AI crawler in robots.txt at all

  2 of them block one

The headline is not the 53.2%. It is that not one of those manifests was authored by the brand. Agent-readiness, at this point in 2026, is something a storefront platform switches on — not something a brand has decided.

How it was measured

• Unauthenticated HTTP requests only. No login, no crawling behind a session — nothing a brand could not reproduce against its own storefront.
• One sampled product page per storefront for the PDP schema columns, not an exhaustive catalogue scan.
• Issuer attribution is the part we redid. The original collection labelled a manifest "Shopify" only when the host hint leaked a *.myshopify.com address — that is 3 of 25, and it under-reports badly, because a Shopify store on a custom domain leaks nothing. On 2026-08-21 we re-requested all 25 manifests and read the powered-by response header: 22 answered Shopify. The remaining 3 (Anker, Soundcore, eufy) no longer answer at that path at all, but their collection-time host hints were *.myshopify.com. That is 25 of 25.
• Percentages use the 47 reachable storefronts as the denominator, not 50. Three did not answer.

Limits — read these before citing

• n = 50, purposively sampled from cross-border DTC brands selling into the US. Not a random sample; do not project it onto DTC generally.
• A single point in time. An earlier version of this page reported that three manifests had gone dark between collection and re-verification (~12% churn). That was withdrawn in v1.0.2: the collector requested /.well-known/ucp while the re-verification script requested /.well-known/ucp.json, and self-hosted frontends answer only the first — so the three were false negatives. On re-test all 25 manifests were still served.
• What the corrected re-test did surface is sharper: 22 of the 25 manifests were rewritten between April and August 2026 (version 2026-04-08 → 2026-08-25), while the three self-hosted frontends stayed on the old version. A panel that tracks only whether a file exists would report that month as “nothing happened”. Existence is the wrong instrument; version and content are.
• Presence of a file is not quality of a file. We did not evaluate whether the manifests describe the catalogue well.
• This measures what a storefront exposes. It does not measure whether any AI agent actually used it — nobody outside the engines can measure that today, and we will not claim otherwise.

Fields