New AI Fixes Shopping Recommendations by Teaching Itself From Mistakes

If you've ever wondered why your shopping app keeps suggesting things you already own, this research is for you. A team of computer scientists built MERIT, a new AI framework that helps e-commerce sites understand your interests from hundreds of thousands of possible categories. The big idea: instead of only training AI on perfect examples, they taught it to recover from its own mistakes. That small change made a big difference in how useful recommendations feel.

The technical problem has a name: exposure bias. During training, recommendation AI usually gets the correct answer at every step — like a student taking a test while holding the answer key. But in the real world, there is no answer key. The AI makes one small guess, and if that guess is wrong, it affects every guess after it. That's why you see the same type of product over and over, while other things you'd actually like never show up. MERIT fixes this by mixing correct answers with deliberately wrong ones during training, forcing the model to learn how to correct course when it slips up.

On a private e-commerce dataset with more than 250,000 interest categories, MERIT improved global recall — how many true interests the system finds — by at least 11.9%. It also improved Hit@k, the chance that something you genuinely want lands in your top recommendations, by 6.1%. Even better, in live A/B tests with real shoppers, it lifted user conversion by 0.26%, meaning more people actually bought something.

A 0.26% lift sounds tiny, but at the scale of a large online store, that can mean millions in extra revenue. The catch: the research uses proprietary data, so we don't yet know how well it works across every platform. Still, it points toward a future where shopping AI is less repetitive, more adventurous, and better at finding you things you didn't know you wanted.