Meituan's New AI Shows You Ads That Actually Fit Your Shopping

When you scroll through an app like Meituan — China's giant platform for food delivery, shopping, and local services — you see ads mixed in with the search results. These ads are chosen by a "click-through rate" (CTR) prediction system: software that guesses how likely you are to tap on each ad. The better the guess, the more money the platform makes and the more useful the ads feel to you.

For years, these systems treated your past clicks and your current search as two different things. UniCon, a new model from Meituan's researchers, changes that. It treats everything you do in one shopping session as a single thread — one continuous context. Instead of looking at isolated clicks, it looks at the whole flow: you searched for milk, then looked at cereal, then browsed a coupon page. Each of those stops is part of one story, and UniCon uses that story to predict what you want next.

This approach pays off in real numbers. When Meituan ran it on actual search ads, ad revenue jumped 2.95%, clicks rose 2.07%, and overall income per viewer went up 3.09%. That may not sound massive, but for a platform serving millions of people daily, even a 2% lift is worth a fortune. The researchers also say the system is faster and cheaper to run, which means these gains come without slowing down your app.

The catch: this only helps inside platforms that adopt it, and it still exists to sell you more. But a smarter prediction system doesn't just benefit the seller — it can also surface deals and services you genuinely want, cutting down on spammy, irrelevant ads.