Kurly boosts conversion rate by about 100% with cart complementary product recommendations

Kurly's Data Service Development Team applied a complementary-product-based personalized recommendation model through a bottom sheet recommendation feature introduced between the cart and checkout pages. Drawing on behavioral economics research showing that at Stage 2 (purchase decision) of the e-commerce purchase decision journey—the cart stage—recommending complementary products rather than substitute products increases willingness to pay (WTP), the team designed the feature accordingly.

Model Selection and Training Strategy The initial BERT4Rec model suffered from a problem where certain popular products (e.g., milk) were excessively recommended. To address this, the team used NPMI (Normalized Pointwise Mutual Information) to quantify complementary relationships between categories, and trained the model by splitting order data into subsets according to complementary category pairs. In addition, to ensure diversity in recommendation results, they implemented logic that randomizes the spacing between products of the same category, drawing reference from Spotify's Shuffle algorithm.

A/B Test Results An A/B test was conducted over 1 week targeting about 8% of all users. Compared to the control group, which received non-personalized recommendations, the experimental group (BERT4Rec-based complementary product recommendations) saw the proportion of users who converted in the cart, the number of products added per user, and the amount of products added per user all rise by nearly 100%. This increase was confirmed to be statistically significant, and the model is currently being served to a specific user segment in the Kurly app.