Alibaba's New AI Search Predicts What You'll Buy on Tmall

Alibaba researchers unveiled a new way to power search on Tmall, one of China's biggest shopping apps. Traditional search works like a librarian walking the aisles: you type something, and the system looks up products that match. The new system, called VARG, flips that. It uses AI to generate the product IDs directly — as if the librarian simply hands you the right items without searching the shelves. The team says this makes results faster and more personalized.

What makes VARG unusual is that it doesn't just chase relevance. It also bakes in business value from the start, meaning the system is trained to favor products that make the company more money. Each product gets a short coded "address," and part of that code quietly signals how valuable the item is to the business. The AI is then fine-tuned in stages: first learning to match products to those codes, then learning to understand what your search words mean, and finally learning your personal habits.

To test it, Alibaba ran a 14-day experiment covering 20% of real Tmall search traffic. The results were modest but real: total sales rose 1.45%, the number of product pages users viewed per person ticked up 0.22%, and the share of people clicking products rose 0.31%. The system also held its own on shopping-guide style searches while showing fewer candidate products — meaning less computing power needed to do the same job.

The company also solved a practical headache: keeping product codes stable while new items arrive daily. If codes shift every day, the AI's memory breaks. Their update routine adds new products without scrambling old addresses. The catch for shoppers: when a search engine is trained to maximize business value, the top result isn't always the best product — it's the best product for the company's bottom line.