NewPlatformDTC is in public beta
Back to Research
MethodologyAugust 20265 min read

llms.txt Is Not a Proxy for Agent Readiness

PlatformDTC Research

Abstract

Most current advice assumes a store's llms.txt indicates how prepared it is for agents. Requesting /llms.txt and the discovery path on the same 58 hosts shows the two are only loosely related: eight stores accept a full agent checkout while publishing an llms.txt that never mentions it, and four of those publish no llms.txt at all.

Key Findings

1

8 stores accept a full agent checkout while their llms.txt says nothing about it

2

4 of those 8 serve no llms.txt at all; one serves HTML under a 200 status rather than a text file

3

Roughly a quarter of agent-transactable stores in the sample are invisible to the audit method the category is using

4

Each row was confirmed twice — a versioned merchant profile at HTTP 200 with registered payment handlers, and the llms.txt response classified by content-type rather than status code

The Assumption Under the Advice

Audits are built on llms.txt. Adoption percentages are quoted from it, including by this team. The implicit model is that a store which has thought about agents will say so in that file, and one that has not will be silent. Requesting both paths against the same hosts tests the model directly.

The Model Does Not Hold

Eight stores in the sample accept a full agent checkout while publishing an llms.txt that never mentions the endpoints their store answers on. Four of them publish no llms.txt whatsoever. One returns HTTP 200 with an HTML body, which any audit counting status codes rather than content types will record as a successful file. That is roughly a quarter of every agent-transactable store in the sample rendered invisible to the standard method.

Measuring by Content-Type, Not Status

The methodological point generalises beyond this study. A soft-404 — an HTML error page served under a 200 — is the single most common way an automated audit records a file that does not exist. Classification here is by what the host actually served, and the difference between that and status-code counting is large enough to change the headline.

Conclusion

A store's own declaration is not a reliable measure of its capability, in either direction. Readiness has to be measured by calling the endpoints that would actually be used. Any audit that reads only the declaration file will systematically undercount, and the size of the undercount here is about a quarter.

References

  1. [1]PlatformDTC Research (2026). Eight Stores Accept an Agent Checkout. Their Own llms.txt Never Says So. https://platformdtc.com/blog/llms-txt-hides-agent-checkout