Agent-Commerce Discovery Across 147 Direct-to-Consumer Brands
PlatformDTC Research
Abstract
A replication at 2.5 times the original sample. Seventy of 147 direct-to-consumer brands serve a live agent-commerce discovery document — 48% of the full list and 65% of those that answered a request at all. The headline survived the wider frame; two supporting numbers did not, and both are corrected here.
Key Findings
70 of 147 brands serve a live agent-commerce discovery document (48% of the list, 65% of reachable hosts)
Every one of the 70 reports the same protocol version, and 69 answer a tool-listing request with the same thirteen tools
Discovery rate falls with brand size — 74% on the original list, 59% across the expansion — as a platform-default explanation predicts
39 of 147 domains (27%) refuse or fail to answer an identified crawler, almost all while publishing a permissive robots.txt
Corrections
An earlier 30-domain sample reported that no brand blocks AI answer engines in robots.txt. At 147, two do, and both block the full set. Blocking is rare and highly selective — 2 of 147 for answer engines, 6 of 147 for training crawlers — but it is not zero.
Edge refusal was reported as 7 of 30 domains. At 147 it is 39, just over a quarter. The larger sample moved the absolute substantially while leaving the direction intact.
Why Replicate
A 58-domain sample is small enough that any striking result deserves a second look before anyone builds on it. The original study reported that a majority of well-known direct-to-consumer brands already answer agent-commerce discovery requests. That is a large claim about infrastructure nobody announced, so the sample was widened to 147 brands and the collection was rewritten as a script anyone can execute against the published domain list.
What Held
The central finding survived. Seventy of 147 brands serve a live discovery document. The expansion deliberately reached for categories the first list was thin on, and for brands large enough to run their own infrastructure — exactly the population where a platform default would be least likely to apply. The rate fell from 74% to 59% across that expansion, which is the behaviour the platform-default explanation predicts rather than evidence against it. Uniformity is the strongest signal: identical protocol version across all 70 documents, and an identical thirteen-tool surface on 69 of them. This is one vendor's rollout observed from outside, not 147 independent integrations.
What Did Not Hold
Two supporting numbers moved. The first study found no brand blocking AI answer engines in robots.txt; at 147 domains, two do, and the variation among brands blocking training crawlers shows deliberate choices rather than copied defaults — some block everything in the class, others name two crawlers and leave the rest free. The second correction is larger. Edge refusal rose from 7 of 30 to 39 of 147. Nearly every one of those domains publishes a permissive robots.txt while the edge returns 403. The file says yes and the edge says no, and only one of the two is ever checked.
Limits of the Method
The client identifies itself honestly, and no bot vendor recognises the name. A 403 therefore proves that a site refuses unfamiliar crawlers. It does not prove the site refuses a crawler arriving with a name operators know and IP ranges they can verify. Twenty-seven per cent is an upper bound on edge blocking rather than a measurement of it, and it is stated as one. The same caution applies to discovery: absence of a document at a customer domain does not establish absence of the capability at the platform, as headless front-ends in the sample demonstrate.
Conclusion
Agentic commerce is discussed almost entirely in terms of intent — which platforms say they will support which protocol, and when. This corpus takes the opposite approach and calls the endpoints. The result is that a large share of the category is already transactable by an agent, mostly without the merchant's knowledge, and that the audit methods being used to measure readiness miss a quarter of it. The collection is a script, so both the finding and the corrections can be re-derived rather than taken on trust.
Published In Full
Read the complete study, method and datasetReferences
- [1]PlatformDTC Research (2026). We Re-Ran the Study on 147 Brands. Two of Our Numbers Were Wrong. https://platformdtc.com/blog/147-brands-agent-commerce
- [2]PlatformDTC Research (2026). We Checked 30 DTC Brands' robots.txt. None Block AI Answer Engines. https://platformdtc.com/blog/dtc-brands-blocking-ai-crawlers