Agent-ready ≠ agent-working: 25% of "agent-ready" Shopify stores silently fail at add-to-cart

TL;DR: I ran an automated buyer against 20 Shopify stores that explicitly advertise themselves as ready for AI agents (verified UCP merchants). Only 65% let the buyer reach checkout. 25% silently broke at add-to-cart — the button was clicked, the cart stayed empty. A store can declare itself "agent-ready" and still be broken for the agents that actually matter.

The numbers

The finding: there's a gap between declaring readiness and being ready

Every one of these 20 stores has a UCP manifest — they went to the trouble of publishing the file that tells AI shopping agents (ChatGPT, Perplexity, Google's agentic checkout) "you can buy from me."

The existing tooling (UCP Checker, Cloudflare's Agent Readiness score, Shopify's own scanner) checks whether that manifest exists and is well-formed. It does not check whether the store actually works for a non-human buyer.

I checked the thing nobody else checks: does the purchase actually complete?

• 2 stores (dixxon.com, aeternum-jewelry.com) accepted the "Add to cart" click, then served "your cart is empty." The add silently failed.
• 1 store (extoolsjapan.com) showed a permanently-disabled add button with no selectable variant exposed to automation.
• 1 store (fillhappy-va.com) had no add-to-cart at all — its "product" was actually an event page.

That's a real, measurable gap: declared agent-readiness and actual agent-workability diverge, and the divergence is silent — the store thinks it's fine, the agent just doesn't buy.

Methodology (reproducible)

• 20 stores, all from the public UCP registry (ucpregistry.com), all verified, all Shopify.
• An automated buyer: headless Chromium + a polite, identified user agent.
• Per store: found a real product → selected a variant if present → clicked add-to-cart → verified actual cart state (navigated to /cart, checked for a line item) → attempted checkout.
• No payment was ever submitted. "Reached checkout" = the email/payment fields rendered.
• Full per-store results and the harness are public: results.json and checkout_test.py.

Honest caveats

1. This is a scripted browser, not an LLM agent. Real AI agents drive browsers the same way but are less reliable than a script — a silent failure that breaks a script will break an agent too, probably more often. Treat 65% as an optimistic ceiling for real agent success.
2. "Reached checkout" ≠ "purchased." Payment, 3DS, and fraud checks are the next failure layer and weren't tested. Real purchase completion is lower than 65%.
3. Small, biased sample. 20 stores, one platform, and the cohort is self-selected (these stores want agent traffic). The average merchant is almost certainly worse.
4. 3 of the 5 add-to-cart failures are ambiguous (cart state couldn't be verified — could be my detector, not the store). 2 are definitive ("your cart is empty" after a successful add click).

Why this matters

If you're a merchant: agentic checkout is coming (Google's UCP, Perplexity's Buy with Pro, Visa's agent payments are all live). A manifest doesn't make you ready — an agent actually completing a purchase does. The stores losing out are the ones whose carts silently break.

If you build in this space: stop scoring manifests, start scoring purchases. The manifest tells you what a store says; the cart tells you what it does.

Want your store checked?

If you want to know whether your store actually works for automated/AI buyers — not just whether its manifest is valid — email mythrilsredraw@gmail.com. I'll run the same test against your storefront and tell you exactly where it breaks.