The usual framing of llms.txt is that it is a cheap bet nobody has taken yet: no major assistant has publicly committed to consuming it, so publishing one is a small wager on a convention that might matter later. We published that framing ourselves this week.
It is wrong, and the data is not close. We checked 58 well-known DTC brands. 73% of the reachable ones already have an llms.txt — and 70% of those did not write it.
What we found
| Result | Brands | Share |
|---|---|---|
| Has an llms.txt | 33 | 73% |
| — of those, a platform-generated template | 23 | 70% of adopters |
| — of those, bespoke or another format | 10 | 30% of adopters |
| No llms.txt | 12 | 27% |
Of 58 domains, 45 answered an identified crawler; 13 refused or failed to connect. Among the 45, thirty-three serve a real llms.txt — a text or markdown response with actual content, not an HTML soft-404.
Twenty-three of them are the same file
The clue was the file sizes. Twenty-three of the thirty-three cluster between 4.2KB and 4.4KB, and every one opens with the same line:
# Agent Instructions — <Brand>
This document describes how AI agents can interact with
<Brand>'s online store at https://<brand>.com.We checked what platform those brands run on. Ten out of ten sampled were Shopify. The bespoke files were mixed — Warby Parker, Purple and Blue Apron are not Shopify; Outdoor Voices and Burrow are Shopify stores that wrote their own anyway.
So the adoption number is real and the intent behind it mostly is not. Allbirds, Glossier, Casper, Ritual, Brooklinen, Fenty Beauty, Gymshark, Everlane and Kosas all publish an llms.txt. As far as we can tell, none of them decided to.
Then we read the template, and it is not boilerplate
We had assumed the generated file was a polite description of a generic store. It is not. It is a protocol handshake, and it points at endpoints that work.
The template tells an agent that the store implements the Universal Commerce Protocol, gives it a discovery URL at /.well-known/ucp and an MCP endpoint at /api/ucp/mcp, lists supported UCP versions, sets out a six-step agent flow from discovery to completed checkout, and states a hard rule that checkout requires contemporaneous human approval.
So we called them. Every store we tried returns a live UCP merchant profile — version 2026-04-08, with declared services, capabilities and payment handlers including Google Pay registered per merchant. And tools/list against the MCP endpoint returns 13 tools on every store we checked — Allbirds, Glossier, Gymshark, Everlane, Casper and Kosas — covering the full checkout lifecycle: create_checkout, update_checkout, complete_checkout, cancel_checkout, get_cart and more, each with a full JSON schema.
That is not a roadmap or an announcement. It is a working agentic commerce API, deployed per merchant, answering requests today. The merchants did not build it and mostly do not know it is there — which makes the llms.txt less an adoption statistic than the place the platform announces what it already did on their behalf.
It also reframes the “which platforms are agent-ready” conversation rather completely. The listicles that grade platforms on native UCP support are grading a capability that one platform has quietly shipped to its entire base, live, with agent payment handlers attached.
What this changes
Three things, and the first is about our own advice.
Publishing one is no longer a differentiator. We described it as a cheap asymmetric bet. On this evidence it is table stakes — if you are on Shopify you almost certainly already have one, and the “should we bother” debate is settled by your platform rather than by you.
The default is doing more than describing you. We expected the generated file to be a generic summary and to argue that a bespoke one is the differentiator. Having read it, that is only half right. The template carries a working protocol integration, which is worth considerably more than prose. What it does not carry is anything about which of your products matter or what your buyers actually ask — so a bespoke file is a complement to it, not a replacement. The brands that wrote their own range from 566 bytes to 28.5KB.
The convention has quietly won on adoption. Whether assistants consume llms.txt is still genuinely unsettled. But the argument that nobody publishes one is now empirically false, and a standard that ships by default on a major platform has a very different trajectory from one that requires every merchant to opt in.
Why this differs from the retail numbers
The closest existing dataset is a June 2026 snapshot of 15 large retail domains — Amazon, eBay, Nike, Walmart, Target, Best Buy, Home Depot, Wayfair, IKEA, Nordstrom, Macy’s and others. It found 3 of 15 publishing an llms.txt, roughly 20%, against our 73%.
That gap is not a contradiction; it is the same finding seen from the other side. Those are enterprise retailers on Adobe Commerce, Salesforce and bespoke stacks. Almost none of them run Shopify, and Shopify is what generates the file. If adoption were driven by merchants deciding llms.txt was worthwhile, the two samples should look similar. They do not, and the variable that separates them is platform rather than intent.
The same snapshot found only Amazon and eBay restricting AI crawlers, a 16.7% block rate, with every other sampled retailer allowing all tracked crawlers. That is consistent with our own robots.txt study of 30 DTC brands, where none blocked answer engines — and with a 2025 peer-reviewed study of 40,455 popular websites, which found roughly 8–10% fully blocking at least one AI crawler, rising to 12–14% among the largest sites, with GPTBot and CCBot most frequently restricted.
Three samples, three methods, one conclusion: the ecommerce blocking panic is not supported by what sites actually publish. What varies is how much of the visible behaviour is a choice.
What this does not say
Our sample skews to Shopify, which is exactly the population where the effect appears — a sample of enterprise brands on Adobe Commerce or Salesforce would very likely show far lower adoption. Read the 73% as a statement about well-known DTC brands, not about ecommerce.
We also did not evaluate quality beyond the template/bespoke split, and we did not verify that the templated files are accurate about the stores they describe. And thirteen domains refused an identified crawler outright, which means our denominator excludes exactly the sites most likely to be restrictive about machine access.
Method
One request each to /robots.txt and /llms.txt per domain, 14 August 2026, redirects followed, user-agent PlatformDTC-AgentReadiness/1.0 with a contact URL. A response counted as a real llms.txt only if it returned 200 with more than 20 bytes and a non-HTML content type, which excludes SPA soft-404s. Template classification is a match on the fixed opening line above. Platform detection is a check for Shopify CDN and theme markers on the homepage.
One domain was dropped from the source list after checking: merit.com resolves to Merit Medical Systems, not the beauty brand, so its 63KB llms.txt is a domain collision rather than a data point.
Our own llms.txt is at platformdtc.com/llms.txt, generated from our route registry so it cannot rot, with the full corpus at llms-full.txt.
The data
Every figure above comes from a published, per-brand dataset you can check rather than take on trust: the DTC AI crawler index. Free to cite with attribution. Related reading: We checked 30 DTC brands’ robots.txt.