Duolingo uses a machine learning model to decide, in real time, whether to show each individual user an advertisement, a subscription offer, or neither — predicting which choice is most likely to generate revenue from that specific person at that moment. The system was credited with generating tens of millions of dollars in additional annual revenue.

Disclosed in a March 2025 blog post, the system was built on dbt and BigQuery ML infrastructure using an XGBoost model configured as a contextual bandit — an AI approach where the system learns over time which decisions generate the best outcomes. The key innovation was shifting from predicting baseline purchase probability to modeling causal uplift: estimating how showing a particular message actually changes a user's likelihood of subscribing, rather than simply identifying users who were likely to subscribe anyway. This distinction allowed Duolingo to target persuadable users rather than those who would have converted regardless. The system was cited as accounting for approximately 25% of year-over-year revenue growth.