Conversion rate

Meta’s research distinguishes conversion-rate prediction tasks labeled post-impression offsite CVR, post-click offsite CVR, MAI CVR and inline CVR2. Meta’s research reports different results for these tasks, including “NE diff (%)” values of −0.09 for post-impression offsite predictions and −0.05 for post-click offsite predictions2.

These individual task labels differ from the joint task grouping in Meta’s Lattice research, which combines click-rate and conversion-rate prediction across different attribution windows into three tasks for its internal dataset12. Meta’s research uses shared models to learn from these tasks together, aiming to provide a foundation that can benefit more specialized prediction tasks1. For these internal tasks, Meta’s research evaluates normalized entropy against a baseline model used as the comparison point1.

To balance prediction goals, Meta’s MultiBalance research examines a trade-off between conversion-rate prediction and other advertising predictions3. In Meta’s ads-model experiment, increasing the task-weight learning rate—how quickly the model adjusts the importance assigned to each prediction goal—improved conversion-rate prediction performance but hurt predictions of how often ads receive clicks and how often ad displays lead to conversions3. Meta’s researchers selected a task-weight learning rate of 1.0 after examining this trade-off3.

Where MultiBalance studies the balance between prediction goals, Meta’s engineering team describes modeling thousands of user event sequences, including clicks, views and purchases, to build behavioral representations that improve ad relevance and conversion rates across its Family of Apps34. Meta’s engineering team reports that these representations, together with broader modeling innovations, drove a cumulative 6% lift in conversions on Instagram4.