New AI Pretends to Be a Shopper to Make Online Shopping Smarter

Online stores constantly tweak what they show you — search results, product recommendations, ads. But testing those changes on real customers is slow and risky. So researchers try to build AI "shoppers" that act like you do, using them to test new features before going live. The problem: earlier AI shoppers were bad at remembering the longer journey. You might look at ten pages over a shopping trip, but the AI would forget what you saw early on or get confused by too much info. That made its behavior unrealistic — for example, clicking around too much or giving up too easily.

RecVerse, from a team in China and Singapore, solves this by giving the AI a memory system inspired by how people actually think. It uses three layers: a short-term working memory for what you're looking at right now, an episodic memory for what happened earlier in this shopping session, and a preference memory for your overall tastes and goals. Each time the AI decides to store or update a memory, that counts as an action. So it learns by itself when to remember something and when to let it go — kind of like how you keep an eye on a price while also remembering you wanted a gift for a friend.

The other big change is how the AI learns. Older systems were trained to copy each individual click in a logged session, which sounds smart but actually creates weird patterns. RecVerse instead gets scored on the entire shopping trip, checking that the mix of actions — browsing, searching, comparing, buying — lines up with real human behavior, and that the shopper's underlying intent stays consistent throughout. In tests, RecVerse outperformed other simulators at acting like a real person, and the team released a public benchmark called USB to help other researchers build on the work.

Why should you care? This kind of research is what makes future online stores better. If an AI can faithfully simulate real shoppers, companies can test recommendation systems and website designs without spying on actual customers or running messy live experiments. That could mean fewer irrelevant suggestions, less time wasted searching, and a smoother shopping experience. The catch: a simulation is still a simulation. Real people have moods, sudden urges, and messy behavior that an AI may never fully capture. But for now, it's a clever step toward making bots that finally understand how we actually browse.