ReportOct 28 – Nov 10, 2026

Mystery Shopping with Agents

Agents aren't announcing themselves

KYA — Know Your Agent — assumes an assistant will present an identity credential at checkout. On the demand side, that isn't happening.

01Conclusion

The demand side isn't opting in

Across 7 assistants and 35 store visits (Oct 28 – Nov 10, 2026), no agent presented an identity credential. Stores recognized an agent only about once. Alo built a dedicated agent lane — and nobody used it.

0 / 35visits with an agent credential
1 / 35visits where a store likely spotted an agent
0agents that used Alo's agent lane

02What this means

What this means

For stores

Don't assume an agent will announce itself. Publishing an agents.md or an agent checkout lane is not enough if assistants still arrive as ordinary browsers. Bot defenses also hit plain HTTP harder than a real browser, so a quiet Chrome session can look like a customer.

For KYA providers

The credential only works if the agent presents it. These runs point to an adoption gap on the demand side, not just missing rails at the merchant. Until assistants opt in, a store cannot reliably "know the agent."

For agent builders

Presenting identity — Web Bot Auth, UCP, or a KYA credential — is how a store can treat you as an agent instead of a silent browser. Without that, you inherit bot walls, robots.txt conflicts, and no legitimate agent checkout path.

03Evidence

How far they got

Furthest step at each store. Every run stopped before payment, as the prompt required. Instinct's Nov 10 rerun covers all five stores.

robots.txt onlySearch / blocked / sign-inProduct / priceCartCheckout infoPayment / guest checkout

Compact view of the same 35 visits: whether a credential was shown, whether the store appeared to know, and where the run stopped.

AssistantStoreCredentialStore knew?Furthest step

04Method

How the test was run

We ran one standard prompt in each assistant. The assistant had to say how it handles payment authorization and bot blocks, then try to check out on each store without paying, creating an account, or entering personal details. It also reported whether it read robots.txt and llms.txt, how it identified itself, and how far it got.

  • Instinct reran all 5 stores with the full prompt on Nov 10, 2026, replacing its Oct 28 retail results. A supplemental Macy's pass used a true single pair; both Macy's passes reached guest checkout. It used a browser plus plain HTTPS for public files (plain HTTPS got 403 at Macy's and a 307 block page at Walmart; the browser was not blocked).
  • Muse reran all 5 stores with the full prompt on Nov 9, 2026, replacing its Oct 28 retail results. It called itself "Protein" and reported Meta as its maker and "Muse Spark 1.3" as its model (self-reported, not independently verified). It inferred some policy-file status codes from rendered pages rather than raw HTTP headers.
  • Hark used a shorter 3-question prompt: Macy's, Alo, Etsy, Walmart, and Gymshark (Oct 29, 2026).
  • Pally used the full prompt on all 5 retail stores (Oct 29, 2026).
  • Tomo used the full latest prompt on all 5 retail stores (Nov 9, 2026). It read robots.txt only after each attempt.
  • Caddy used the full prompt on all 5 retail stores (Nov 9, 2026; pasted in parts). It mostly browsed and searched and never added anything to a cart.
  • Grok Bot used the full prompt on all 5 retail stores (Nov 9, 2026). It drove its own Chrome browser and used curl only to check policy files.
  • Assistants with no results are left out.

05Full data

Want the rest?

The full per-visit notes, policy-file reads, errors, and raw logs are available on request. They are not published on this page.