Nearly half of shoppers who bought an AI pick sent it back. CPG brands should fix the data

Blue Yonder's fourth annual Consumer Retail Returns Survey, fielded in September 2026 among more than 5,000 shoppers (2,000 of them in the United States), puts a number on something brand teams already see in customer service queues. Of the 24% of U.S. consumers who have bought a product after an AI tool or assistant recommended it, 46% have returned one of those items in the past year.

The most common reason is not size. It is that the product was different than described. That is a product-data problem, and CPG brands can fix it before holiday boxes start coming back.

What did Blue Yonder find about AI shopping returns?#

The survey asked about AI tools or assistants, such as consumer chatbots, not about a single retailer app. Blue Yonder published the results on October 27. Retail Dive picked them up on November 3. Figures below are U.S. unless noted.

Nearly a quarter of U.S. respondents, 24%, have purchased a product after receiving an AI recommendation. Younger shoppers are doing it far more often: 43% of Gen Z and 36% of millennials, against 19% of Gen X and 9% of baby boomers.

Buying is only half the story. Among people who did buy on an AI recommendation, 46% sent at least one of those items back in the past year. For Gen Z AI shoppers, that share rises to 56%, according to Blue Yonder's press release. Of the people who returned an AI-recommended purchase, 49% said they are more likely to return those items than purchases made directly through a retailer, and only 9% said they are less likely to.

Trust is thin even before the box is opened. Only 20% of U.S. respondents would trust an AI tool to recommend the right size or fit. That figure is 36% among Gen Z, 27% among millennials, 18% among Gen X and 9% among baby boomers.

Why do shoppers return AI-recommended products?#

Among U.S. shoppers who returned an AI-recommended purchase, 36% said it was different than described, and 23% said it was the wrong size or fit. For U.S. returns overall, those reasons flip. Incorrect size or fit leads at 42%, and "different than described" accounts for only 20%.

Blue Yonder's own read is that AI tools struggle less with fit than with describing products accurately. Duncan Angove, the company's CEO, said: "That's not really an AI problem. It's a data problem. An AI assistant is only as good as the product information it can find, and most product data was written for a search bar, not for a machine suggesting products for someone. When what arrives doesn't match what the shopper expected, it goes back in the box."

A detergent listing written to rank for "best laundry detergent" often never says unscented, 150 oz or 64 loads near the top. A person can skim a photo and a bullet list and figure it out. An assistant summarizes the text it can parse and tells the shopper this is a fresh-scent everyday detergent. The shopper who asked for unscented gets a bottle that smells like a meadow. That is a "different than described" return, and the description they heard was the agent's, built from your copy.

The same gap shows up when pack counts, concentrate levels, age limits and diet flags live in an image or a file the agent never reads. We wrote last week about Shopify's finding that AI shoppers reward specific product data. This survey is the other side of that coin. When the data is vague, the recommendation can still happen. Then the product comes back.

How should CPG brands write product data for AI shopping assistants?#

Write the facts an assistant would repeat out loud, and put them where a machine can find them. The title, the first sentence and the structured attributes should name the product type, the size or pack count, and the variant a shopper would say: fragrance-free, 2x concentrate, peanut-free, 12-count.

Say what the product is not. If it is not a sunscreen, not a treatment, not safe for pets or not a single-use bottle, write that in text. Agents fill gaps. Empty gaps become over-promises, and over-promises become returns.

Then check whether those facts survive the first pass. Paste a listing into the tester below. If the claim that would have stopped a "not as described" return sits after the first 40 words, rewrite so it does not.

Do this on the live Amazon, Walmart, Target and brand-site listings. The pages have to agree. An agent that reads two different pack sizes, or a bold claim on the brand site and silence on the retailer page, will invent a middle version. That middle version is what shows up at the door.

We would run that 20-question check every week through the holidays, If an assistant describes your detergent as scented when it is not, that is a listing bug, not a model mystery. A free audit is a faster way to see which SKUs agents name and how they talk about them.

What happens if retailers just tighten return policies?#

The other half of the survey warns against making returns harder and calling the job done. Fifty-one percent of U.S. respondents have chosen one retailer over another purely because of a better returns policy. Twenty-nine percent stopped shopping at a retailer that tightened its policy in the last year, including 47% of Gen Z and 40% of millennials.

If a retailer told them they return too often and limited their returns, 82% said they would cut back or walk away: 43% would stop shopping there, and 39% would shop there less. Sixty-three percent of U.S. consumers would not pay any fee to return a product, the highest share of the five countries surveyed.

Retailers will still try to contain cost. Brands should not wait for that argument. Every "different than described" return that started with an AI recommendation is a SKU whose public data taught the agent the wrong story. Fixing that data protects the shopper, the retailer relationship and the next recommendation.

Our take#

AI shopping is already sending products home, and the returns are concentrated among the shoppers who use it most. The fix is unglamorous: accurate titles, structured claims, matching copy across every store an agent can read, and a weekly check of how those agents describe you. That is the work SearchPanel is built for. If you want help seeing the gaps, run a free audit or book a demo.