Tencent's Ad AI Study: Your Profile Beats Your Browsing History

Every time you click an online ad, a quiet piece of software makes a bet: will this person actually buy something? That bet is called CVR prediction (guessing whether a click turns into a purchase), and it decides which ads you see and how much advertisers pay. In a 2026 Tencent contest, teams competed on 34.82 million real ad records. A pair of researchers finished 10th, and their write-up is more interesting than the ranking.

Their big surprise was about what actually improved the AI. The model was fed two kinds of information: rich profile-style facts about the shopper and the product (age, region, past purchase history, price) and the ordered trail of what the person did (this click, then that one). Conventional wisdom says the behavior trail should matter most. It didn't. Removing the rich profile data hurt the score roughly 20 times more than removing every behavior-tracking component combined. In plain terms: knowing who you are beat knowing what you just did.

The second finding is a warning, and it applies far beyond ad tech. The team found that their internal testing setup — which mixed records from a single time period — made their model look about 0.014 points better than it really was on the live leaderboard. Worse, some fixes that helped in testing actually hurt in the real world, a sign the model had memorized quirks of one data batch rather than learning anything lasting. The lesson: if you test AI on data that looks too much like its training data, you're grading your own homework.

For everyday readers, this matters because CVR-style models decide which ads chase you, which products get recommended, and what prices you're shown. The study suggests companies can get better results — and spend less computing power — by collecting solid profile data rather than endlessly tracking every tap. It's also a reminder to treat flashy AI performance numbers with healthy skepticism until someone verifies them on fresh, real-world data.