AI VK engineers built a separate product recommendation system within the VK feed. The usual post selection mechanism focuses on likes and comments, but interest in a publication does not yet mean that a person is ready to pay. For products, a separate tool was created with its own model that decides what to show and in what order.
At the start, a trap arose: few impressions — few orders — nothing to train the model on. The team increased the recording of user actions on products from 2% to 100% — the training sample grew 15-fold. While orders were insufficient, the system relied on indirect signals, such as expanding the post text: this action often preceded a transition to the seller. Priorities for display are structured as follows: first an order, then a transition to a partner.
Over five months, the share of impressions that resulted in an order increased by 53%. Orders themselves increased 43.2 times, transitions to the seller by 55.3 times, and the click-through rate of these transitions increased by 95.6%. Product posts are mixed into the feed with a probability of 1.5%, and for promising products — up to 5%. Everything runs on the internal VK Discovery platform.