In-depth analysis · About 5 min read
An a16z analysis examines why AI shopping still meets resistance
Moses Sternstein separates product research from delegated checkout, with explicit limits on the underlying evidence.
- Source published
- Published here
- Source checked
In an October 16 a16z newsletter, Moses Sternstein examines obstacles to agentic commerce. His argument concerns the gap between AI helping people research products and agents being accepted or trusted to complete purchases; it is an attributed analysis.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- Sternstein discusses retailer restrictions, customer relationships, returns and disputed transactions as possible sources of friction. These are explanations he considers, rather than evidence establishing the reason for each blocked agent request.
- The article explicitly says some cited shopping clickstream data ends in June. It treats those observations cautiously and leaves open whether the same patterns still hold.
The argument is about permission, not just capability
Moses Sternstein's October 16 newsletter asks why technically capable shopping agents still meet resistance. Its opening section combines access statistics, returns and fraud figures, survey responses and shopping clickstreams. These are third-party observations discussed by an a16z author, not a new controlled experiment conducted by a16z or AISubHub. The useful distinction is between finding information, recommending a product and committing a payment. A demonstration that an agent can navigate a page does not settle whether a retailer permits that access, a customer has authorised the transaction, or either party accepts responsibility when an order goes wrong. Those are separate conditions for a usable service.
A rise in forbidden responses cannot identify every cause
The author cites Cloudflare data showing the share of forbidden responses to agents and bots roughly doubling from an earlier level after rising over the summer. He explicitly separates that observation from the question of whether agentic commerce caused it. The category includes more than shopping agents, and a refusal alone does not reveal the merchant's reasoning. The newsletter discusses customer relationships, advertising revenue, fraud, returns and disputed charges as plausible commercial frictions. It also cites NRF figures suggesting return rates and aggregate return costs have risen substantially since 2019. These arguments explain possible incentives; they do not prove why a particular website blocked a particular request.
Consumer assistance is not delegated purchasing
In survey results reproduced by the newsletter, roughly 25% of respondents were comfortable letting AI purchase several kinds of consumer goods, with luxury goods treated differently, while 42% said they did not trust AI to buy for them at all. These are attributed survey results, not universal preferences or a current count of all shoppers. The article does not supply enough survey methodology in its prose to resolve every sampling or wording concern. The distinction still matters operationally: a shopper may welcome price research without authorising payment, substitutions, a recurring charge or a large quantity. AISubHub reads explicit, bounded approval as a design requirement, rather than assuming a useful recommendation implies consent to checkout.
The clickstream evidence has an important time limit
The Similarweb discussion places many observed AI interactions between discovery and purchase, where people research products before returning to conventional checkout. The author warns that these clickstream data appear to end in June. They cannot establish the behaviour of October users, let alone predict a permanent shopping pattern. Observed navigation also cannot automatically identify which recommendation caused a later purchase in a different session. This makes advertising attribution a measurement problem, not proof that AI recommendations have no effect. Our analysis is that teams should label the observation period, distinguish visits from completed orders, and avoid turning a chart about journeys into a claim about autonomous transaction reliability.
What to check before buying an agentic shopping service
For AISubHub readers, the practical checklist starts with the supported merchant workflow and its permissions. Check whether the service only researches, prepares a cart or actually submits an order; who confirms amount, currency, tax and recurring terms; and how cancellation, returns and disputed transactions are handled. A failed or forbidden request should remain a visible limitation, not become a reason to bypass a site's restrictions. These are our editorial recommendations, not features verified in any subscription product. We have not bought goods through an agent or independently audited the cited charts. The broader conclusion is that progress in model capability must be matched by accountable commercial arrangements and customer control before a persuasive demo becomes a dependable purchasing workflow.
Why it matters
For users comparing AI services, a useful distinction is product research versus authorization to spend money. Check which merchant workflows a tool supports and keep purchase decisions and payment permissions explicit.
Limits and availability
This commentary is not a current census of shoppers or an AISubHub experiment. We have not independently validated its third-party charts; it cannot establish universal consumer preferences or prove that agentic commerce caused every reported access refusal.
Sources and evidence
Source published: Oct 16, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.