Discovery has already moved, checkout has not followed
New PYMNTS Intelligence data puts a number on a shift retailers have felt anecdotally for a year: an estimated 49.6 million US adults, close to 19% of the adult population, now begin retail product searches inside an AI assistant like ChatGPT, Claude, or Gemini rather than a traditional search engine. Of that group, roughly 39 million use AI primarily for product discovery and comparison, while another 10.6 million blend AI assistants with conventional search in the same shopping session.
That volume is large enough that retailers can no longer treat AI referral traffic as a novelty to monitor from the sidelines. It behaves like a new channel with its own conversion patterns, its own attribution headaches, and its own budget line, and the retailers building strategy around it are making a specific and consistent choice: let the AI platform do the browsing and comparison work, then bring the shopper back to a domain the retailer controls for the actual transaction, the loyalty enrollment, and every future touchpoint that follows the sale.
The tax on someone else's platform
Josh Friedman at Ulta Beauty put the logic bluntly: there is always a tax for engaging customers on other people's platforms. That tax shows up as lost margin on any commission the platform charges, lost visibility into what the customer actually browsed and abandoned, and lost ability to remarket without going back through an intermediary that controls the relationship. Ulta's posture is to show up where AI assistants surface product recommendations, but to close the loop on channels it owns.
Etsy's Rafe Colburn frames it as a relationship question rather than a purely financial one. He describes a completed purchase on Etsy's own site as the start of a deeper relationship, not one that stays permanently intermediated by a chatbot. That is a meaningful hedge for a marketplace whose entire value proposition rests on repeat visits and seller discovery, both of which are hard to sustain if every future search routes back through someone else's assistant instead of Etsy's own app or site.
Why the infrastructure providers agree
Vince Koh at AWS makes the same case from the infrastructure side: when a purchase closes on the retailer's own site, the retailer maintains a direct relationship with that customer, complete with the purchase history, browsing behavior, and contact permissions that first-party data requires. That data feeds loyalty programs, personalization models, and retention campaigns that an AI-mediated transaction, closed entirely inside a third-party chat interface, would never generate for the retailer at all.
Raina Moskowitz at The Knot points to three decades of accumulated planning data as the reason a wedding-adjacent retailer needs to own the transaction layer, not just the discovery layer. That kind of proprietary dataset, built purchase by purchase over years, is precisely what evaporates if a growing share of transactions route through and terminate inside somebody else's AI assistant rather than a domain the retailer can instrument. It is also the dataset that trains the retailer's own recommendation models, and a model trained on a shrinking slice of actual transactions gets worse at the exact job it exists to do, which compounds the cost of ceding checkout well beyond the margin on any single sale.
This complicates the agentic commerce narrative
The retail industry has spent the last two quarters celebrating AI-referred traffic for converting at higher rates and driving deeper engagement, more pages browsed, more time on site, than traditional channels. That celebration has mostly glossed over where the transaction itself actually happens. The retailers quoted here are drawing a hard line between welcoming AI-driven discovery and ceding the checkout to the platform that generated it, and that line matters more than the overall traffic number.
It also means CTOs building for agentic commerce need to design two separate integration points rather than one. The first is exposing product data cleanly enough that an AI assistant can recommend it accurately, structured feeds, current inventory, accurate pricing, the same rigor teams once reserved for a search engine's product listing ads. The second, and the one getting less attention internally, is making the handoff back to an owned checkout frictionless enough that the customer does not simply complete the purchase inside the assistant instead out of sheer convenience, taking the data and the margin with it.
The build decision this creates
For a CTO or head of ecommerce, this reframes a chunk of the agentic commerce roadmap. Investment in a rich, fast, AI-agent-readable product feed is now table stakes, the same way a mobile-responsive site was table stakes a decade ago. But the return on that investment depends entirely on the retailer's own checkout being fast enough and low-friction enough that a shopper who discovered a product through an assistant does not simply ask the assistant to complete the purchase for them instead.
That is a UX and identity problem as much as an AI problem: one-click account recognition, saved payment methods, and a checkout flow that survives the jump from a chat window to a browser tab without losing the cart contents along the way. Retailers that solve the discovery layer without solving that handoff will find themselves generating traffic and demand for a competitor's data asset instead of their own, having done the hard work of getting found and handed the payoff to whichever platform closed the sale.
