Nearly a quarter of American consumers have now bought something because an AI assistant told them to. Of those, 46 percent returned at least one of those purchases in the past year. Those two numbers, from Blue Yonder's 2026 Consumer Retail Returns Survey, are being read in much of the trade press as a story about artificial intelligence. Our view is that they are a story about product data, and that the retailers who read them the other way will spend the next two years fixing the wrong thing.
The distinction matters because it determines where the money goes. If AI recommendations are the problem, the answer is better models. If product information is the problem, the answer is better attributes, better images, better packaging lists and better measurement of why things come back. The survey evidence points firmly at the second.
What the numbers actually say
The survey, in its fourth year in the United States, polled more than 5,000 consumers across five markets including 2,000 Americans. Among shoppers who returned an AI-recommended purchase, the most common reason was that the item was different than described, cited by 36 percent, followed by incorrect size or fit at 23 percent. In the overall returns population those two reasons reverse: size and fit leads at 42 percent and description mismatch accounts for only 20 percent. AI is doing a relatively good job of getting fit right and a poor job of conveying what the product actually is.
That inversion is the finding worth building a strategy around. The machine is not guessing wrong about the shopper. It is faithfully repeating wrong or incomplete information about the product, because that information was written for a search box and a human scrolling a grid, not for a model synthesizing a recommendation from a dozen structured fields.
Why descriptions fail
Product data in most catalogs was built for retrieval, not for reasoning. Titles are stuffed with keywords, bullet points repeat the title, and the fields that would answer a specific question about material, compatibility, dimensions or what is included in the box are either missing or buried in free text. A search engine tolerates that. An assistant that must answer whether a phone case fits a particular model, or whether a set includes the mounting hardware, either guesses or hedges, and the shopper arrives with an expectation the product cannot meet.
The categories where this bites hardest are the ones where the physical object differs from its representation: apparel where fabric and drape matter, home goods where color and scale matter, electronics accessories where compatibility is a hard constraint, and anything sold as a kit where the contents are ambiguous. Cross-border sellers carry an extra burden, because the buyer cannot fall back on a local store visit to check, and the listing is the only source of truth across a language and a logistics chain.
The policy trap
The intuitive response to rising returns is to tighten the rules, and the survey shows exactly how that ends. Fifty-one percent of American respondents said they have chosen one retailer over another purely because it had a better returns policy, and 29 percent said they stopped shopping entirely at a retailer that tightened its policy in the past year. When retailers flag shoppers as frequent returners and limit their returns, 82 percent said they would shop less often or stop altogether. Twenty-four percent called return-history tracking invasive enough to send them elsewhere, against 30 percent who considered it fair.
Blue Yonder chief executive Duncan Angove puts the operational problem plainly: some of the most frequent returners are also the best customers, and telling the difference between abuse and loyalty requires systems that can set policy per customer rather than per category. Notably, 63 percent of American shoppers said they would refuse to pay any return fee, the highest share among the five markets surveyed.
What merchants should fix first
Three interventions have measurable payoff and do not require a new platform. The first is attribute completeness on the fields that drive returns: material composition, exact dimensions, compatibility ranges, care instructions and, critically, what is not included. The second is image honesty, which means showing scale, texture and true color rather than the most flattering render, and labeling any composite or staged shot so the shopper does not infer that accessories are part of the purchase. The third is return-reason capture at SKU level, so that mismatch complaints route back into the listing that caused them instead of into a monthly report nobody reads.
There is also a cheaper lever that the survey quantifies. Fifty-five percent of American respondents said a retailer had told them to keep an item they wanted to return in the past year, unchanged from 2025, and 67 percent of those said it made them more likely to shop with that retailer again. Store credit is the distinctly American compromise, accepted by 51 percent of respondents against 36 percent across all five markets.
Our view
The industry has spent two years celebrating AI referral traffic, and the traffic is real: assistant-driven visits are growing fast and, by several platform accounts, converting better than conventional search traffic. But referral quality and post-purchase satisfaction are different metrics, and only the second one shows up in the returns line. Retailers that instrument the gap between what the model said and what the customer received will be able to close it. Retailers that treat 46 percent as a technology problem will keep buying models and keep shipping boxes back.
The generational split sharpens the urgency. Forty-three percent of Gen Z and 36 percent of Millennials have bought on an AI recommendation, against 19 percent of Gen X and 9 percent of Baby Boomers, and 56 percent of Gen Z AI shoppers returned something in the past year. The cohort with the least tolerance for friction is also the cohort adopting the channel fastest, which makes product data a conversion problem rather than a back-office one.
