AI referrals and the checkout question
Shoppers now ask ChatGPT and Google’s Gemini which product to buy. Retailers want the sale that follows to close somewhere else: on their own websites. That tension is rewriting product pages, reshaping deals with AI platforms, and setting up a quiet fight over the customer data that drives every future sale.
Key Takeaways
- Juniper Research expects shoppers to spend $8 billion this year after AI agents such as Anthropic’s Claude and Gemini send them to retail sites.
- Adobe Analytics found that 41% of U.S. consumers used generative AI for online shopping in June, and that AI-referred visitors generated 41% higher revenue per visit than shoppers arriving through traditional channels.
- Walmart, Ulta Beauty and Wayfair have rebuilt their product content to rank in chatbot answers, yet most still want the transaction, and the data attached to it, on their own sites.
Rewriting the shelf for machines
Search engines rank pages on keywords and links. Chatbots answer detailed questions instead. A shopper no longer types “running shoes” and scrolls. She describes her arches, her mileage, and her budget, then waits for a recommendation.
That difference forces retailers to rewrite how they describe products and how customers find brands at all. It is the same optimisation scramble that greeted every previous channel, only faster. Anyone tracking the shift toward chatbot-led product discovery has watched the arithmetic change over a single year.
Ulta Beauty has leaned in. “We always want to be wherever our guest is,” said Josh Friedman, the retailer’s head of digital and e-commerce. “Whether it’s Google search, affiliate marketing (or) Facebook, there’s always a tax for engaging customers on other people’s platforms. I don’t think this is any different.”
Ulta says it is seeing “double the conversion and intent” from shoppers who find its products through Gemini and ChatGPT, a figure the company has not published in detail. It is working with Google to fold shopping carts and its Ulta Beauty Rewards loyalty programme into AI-powered shopping inside Gemini. Friedman still said the retailer would prefer customers finish the purchase on Ulta’s own website.
Retailers defend their turf
Executives argue that a retailer knows its customers far better than any general-purpose assistant does. That knowledge is the asset worth protecting.
“When a shopper completes a purchase on a brand’s own site … the retailer maintains a direct relationship with that customer,” said Vince Koh, global head of digital commerce at Amazon Web Services. AWS encourages retail clients, including Tapestry’s Kate Spade, to market through AI platforms while keeping their own websites the <a href="https://bitcomme.com/best-ai-crm-software-2026/” title=”Best AI CRM Software 2026″>best place to actually buy.
Koh gave a concrete example. When a shopper uses AWS-powered search on Kate Spade’s site to look for a $380 suede bucket bag, the company can record which products she viewed, what she was willing to spend, and which styles she skipped. That record feeds the next recommendation and, eventually, loyalty.
Wedding-planning platform The Knot takes the same line. It is optimising its site so vendors appear in ChatGPT results, then working to convince couples to book venues and invitations through the company directly. “LLMs are another place to start, another place to search, another place to plan, but we have three decades of data to help” people coordinate weddings, chief executive Raina Moskowitz said, referring to AI large language models.
The checkout stays home, for now
Consumers appear to agree, at least on where they press “buy”. In March, OpenAI ended Instant Checkout, the tool that let shoppers purchase clothes and other goods inside ChatGPT. The company now concentrates on product discovery and helps merchants and marketplaces such as Etsy route buyers to their own checkouts. That retreat followed a wave of retailer enthusiasm for in-chat purchasing when the feature first launched.
Etsy reports the same pattern. Users who discover products through ChatGPT typically come back to Etsy’s website to finish the order, said Rafe Colburn, chief product and technology officer. “It’s the start of a deeper relationship, and not just one that’s always intermediated by ChatGPT,” Colburn said.
The payments layer keeps moving underneath all of this. Card networks have begun wiring their own rails directly into chatbots, which would let an agent transact at any participating merchant rather than a short enrolled list. Retailers watching that build-out face the same question in a new form: who holds the customer relationship when the agent handles the money.
None of this is the first time a platform has stood between a brand and its buyer. Search engines, marketplaces and social feeds each took a cut and each collected data along the way. What changed is the speed of the handover. Merchants who tried in-chat purchasing reported thin sales and conversion rates well below their own sites, and several concluded that the discovery half of the deal was the valuable half. Retailers have used that reprieve to build dedicated apps inside chatbots, where they control the presentation, the loyalty programme and the payment.
The economics explain the urgency better than any strategy memo. Data covering more than a trillion visits to U.S. retail sites shows AI-referred visits now carrying a substantial revenue premium over traditional channels, a reversal from a year earlier when ordinary traffic was worth far more. A smaller, higher-intent stream of buyers is worth defending. It is also worth optimising for, which is why so many retailers spent the past year auditing whether a language model can even read their product pages. Those that started early with structured product data and machine-readable descriptions are the ones showing up in answers today.
Meta Launches Muse Code, a Terminal Agent for Large Repositories
Meta Joins OpenAI and Anthropic in Reporting an AI Test Breach- ByteDance Founder Tells AI Team to Stop Distilling Rival Models
White House Will Not Safety-Test Open-Weight AI Models (At Least For Now)
Anthropic Builds In-House Chip Team to Run Claude Faster
More news about Information processing
- The Last Year for Cash? EU is Cutting Limits Next Year
- New research enables methanol use in dual fuel marine engines
Researchers Find Hidden Backdoor in Chinese-Made Zbtlink Routers
Sega’s Dreamcast Is Still Getting New Games 25 Years Later- Underground acoustic signals reveal hidden tunnels
