For two years the conversation about AI agents in commerce was theoretical. Demos at conferences, screenshots of a model browsing a site, a research paper from one of the labs. By the fall of 2025, the conversation has changed. Real consumer-facing agents — some from the model providers themselves, some from new entrants, a few stitched together by aggressive aggregators — are now making transactions on behalf of users at a volume retailers can finally see in their logs. And most retailers we spoke to don't have a policy.

The fingerprint problem

The first practical question every operator is asking is how to tell an agent from a human and from a scraper. The answer in late 2025 is unsatisfying: it depends.

Some agents identify themselves through user-agent strings or signed headers. Others route through residential proxies and look exactly like a Chrome session from a suburb of Phoenix. A director of digital at a national apparel retailer described it this way: "We can identify maybe sixty percent of agent traffic confidently. The rest is a guess based on session patterns — no scroll, no mouse movement, no product page dwell, straight to add-to-cart and checkout in under twelve seconds."

The platforms that have published an agent-friendly stance — including a few well-known retailers that have built explicit agent landing pages and authentication flows — are getting cleaner data. The platforms that haven't are getting a mix of agent traffic and conventional bots blurred together.

The pricing question

The more interesting strategic question is whether agent traffic is worth the same as human traffic. The early evidence in 2025 suggests it isn't — at least, not on the same terms.

Agents convert at extremely high rates because they're not browsing; they've already decided. They take advantage of every available coupon, loyalty mechanic, and promo stack. They don't see ads in the conventional sense. And they tend to return at conventional human rates, because the customer behind the agent is still the one who has to deal with the box.

Operators we spoke to are split on what to do about this. Some are quietly suppressing certain promotions for traffic they suspect is agent-driven. Others are building tiered pricing — a base price for agents, a separate price for authenticated logged-in customers. A few are doing the opposite, paying the agent platforms for placement on the theory that being the default answer to a buying query is the new equivalent of being the first organic search result.

The marketplace platforms are moving first

Among the large platforms, the early movers in agentic commerce are predictably the marketplaces. Amazon has both blocked and launched agents this year — blocking some third-party agents that scrape product data while quietly piloting its own agent-style assistant inside the shopping app. Walmart's marketplace has published initial guidance for agent traffic. Shopify has shipped agent-friendly APIs at the merchant level, leaving the policy decision to the brand.

The pattern is that platforms with a marketplace structure tend to lean into agent traffic because their fee model — a commission on the transaction — captures value regardless of how the buying decision was made. Pure brand sites have more to lose, because the agent strips out the merchandising, the cross-sell, and the AOV machinery the brand spent years building.

What changes for retail media

The implication that's starting to register inside ad teams: if agents don't see ads, retail media is selling impressions to a shrinking pool of human eyeballs even as the on-site GMV stays flat or grows. A CPG media buyer told us their late-2025 measurement work is starting to surface this gap, and they expect a hard conversation with retailers heading into 2026 upfronts.

What to watch into Q1 2026

Three things to track: whether any major retailer publishes an explicit agent traffic policy with public terms; whether the model providers agree to a common identification standard (the rumor mill in fall 2025 has at least two competing proposals being shopped); and whether agent-driven returns rates diverge from human ones in a meaningful way once the holiday data lands.

The retailers that come out of Q1 2026 with a clear stance — not necessarily the right one, just a clear one — will have a head start. The ones still treating this as a 2027 problem are going to be making decisions in production logs that they should have made in policy documents.