Product Recommendation Engine Specification Template

Mature e-commerce businesses attribute 15-35% of revenue to recommendations, depending on catalog size and personalization depth. Amazon has famously attributed 35% of its revenue to recommendations. For a new implementation, target 8-12% in year one and grow from there. Track revenue attribution using last-click or assisted-conversion models, but be consistent.

Buy (managed services like Amazon Personalize, Algolia Recommend, or Dynamic Yield) if you have fewer than 500K MAU and limited ML engineering capacity. Build custom if you need deep personalization, have unique data signals, or recommendations are a core differentiator. The build vs. buy decision follows the same framework as any [build vs. buy assessment](/glossary/build-vs-buy). Most teams should start with a managed service and migrate to custom models only when they hit the service's limitations.

Use session-based recommendations: track browse events within the current session (via cookies or session ID) and recommend similar items. This provides relevant suggestions without requiring login. Once a user creates an account, merge session history with their profile. Anonymous users typically represent 60-80% of traffic on e-commerce sites, so session-based recommendations are not optional.

Popularity-based models should update hourly to reflect current trends. Collaborative filtering models should retrain daily or weekly depending on data volume. Content-based models only need retraining when catalog attributes change. Real-time models (session-based) update with every user interaction. The right cadence depends on how quickly your catalog and user behavior change.

Split traffic at the user level, not the session level, to avoid inconsistent experiences. Run each test for at least 2 weeks to capture weekly seasonality. Measure add-to-cart rate and revenue per session, not just CTR. A recommendation change that increases clicks but not purchases is likely recommending clickbait products. Test one variable at a time: algorithm changes and placement changes should be separate experiments. ---