Yandex rolls out a new conversion prediction model for its advertising platform

Yandex has updated the model that predicts conversions and manages advertisers' bids. The new model uses a “multi-head” architecture: several neural networks run inside it, each handling its own type of signal. This lets the system analyze the full user journey on a site rather than isolated actions. Tests showed the model can deliver advertisers an average of 10% more conversions for a comparable budget.

The new model draws on more than 25 signals, three times as many as before. These include pages per visit, time on site, add-to-cart events, and other micro-conversions that point to genuine purchase intent.

Alexey Shtokolov,
Director of Product and AI, Yandex Advertising

There can be a gap between when an ad is shown and when a purchase happens, but advertisers need to judge an ad's performance quickly. The multi-head architecture predicts conversions more accurately because it considers different types of signals at the same time, from a click to a delayed purchase. That means we can learn even from visits to a site or app that didn't immediately lead to a conversion.

Tracking delayed conversions is especially valuable in categories with a long decision cycle, such as real estate, cars, consumer electronics, and travel. A user may spend days or weeks comparing offers and returning to a site before deciding.

Another feature of the new architecture is a built-in proxy model. It picks up recent shifts in audience behavior without lengthy retraining and passes them straight to the main model. With the proxy model in place, the system recognizes intent before the final purchase and adjusts auction bids in real time based on conversion likelihood. As a result, ad budgets go further: ads are shown to the users most likely to buy, fill out a form, call, or download an app.

The new model is live in campaigns on Yandex Search and the Yandex Advertising Network, which includes the company's own services and more than 90,000 partner properties: websites, mobile and messaging apps, blogs, connected TV (CTV), and more.