How should recommendations change when a shopper keeps returning the same category?

October 29, 2026

Short answer: Returns by category are preference data with the noise removed: the shopper kept what worked and sent back what did not. Recommendations should quietly suppress the repeatedly returned pattern, size, or style, and favour the profile of what was kept. Say nothing about it to the shopper. Just get better.

Kept is a stronger signal than clicked

Clicks are curiosity. Purchases are intent. Keeps are truth. A shopper may click everything, buy several things, and keep exactly the subset that fits their body, their home, or their taste. Most recommendation models train on the first two signals and treat the third, the return, as a logistics event in another system.

Joining the systems changes the picture fast. The shopper who returns every pair of slim trousers and keeps every straight cut has filed a fit report more reliable than any size quiz. The system recommending a fourth pair of slim trousers is not being personal. It is being forgetful with confidence.

Suppress the pattern, not the category

The error to avoid is over correction: this shopper returns trousers, therefore hide trousers. Category level suppression punishes a customer for your fit problem and hands the sale to a competitor whose trousers fit. Suppress at the level the evidence supports: the cut, the fabric weight, the brand whose sizing runs small, the heel height band that always comes back.

Reason codes, where shoppers give them, sharpen the level. Too small and runs small are different findings, as are changed mind and not as pictured. A model fed raw return flags without reasons will learn some of this from outcomes, slowly. A model fed the reasons learns it this quarter.

Never announce the inference

The shopper did not ask to be profiled by their returns, and a message saying we noticed you return a lot will land as an accusation no matter how it is worded. Apply the learning silently: the rails change, the emails feature what gets kept, the returned pattern fades. If the shopper notices anything, it should be that the store suggestions got better, full stop.

There is also a fairness line inside the data. Returns driven by damage, late delivery, or your description errors are your faults, and they must be excluded from the shopper pattern before any model sees them. Counting your own mistakes against a customer fit profile is how a personalization project quietly becomes a discrimination complaint.

The metric that tells the truth

Watch kept revenue per recommended session for the shoppers in the pattern, not click through and not gross sales. If the suppression logic works, gross orders from those shoppers may dip while kept revenue rises, and the warehouse stops processing the same trousers in a circle. That trade is the entire point.

Keep reading: Recommendations for gift buyers: a different intent needs a different feed, When to show bestsellers and when to show long tail picks, and Cold start for returning visitors: using the second visit well cover the nearby cases in this series.