Building Recommendation Systems That Drive Revenue
Most recommendation systems are optimized for engagement. Click-through rate, watch time, scroll depth. The implicit assumption is that engagement and revenue are correlated. They often aren't. This piece is about building systems that close that gap.
The engagement trap
Optimizing for engagement maximizes the wrong thing. An algorithm that surfaces the most clickable item gets you to one interaction. An algorithm that surfaces the item most likely to convert gets you a transaction, a repeat customer, and a positive unit economic outcome. The difference is intent modeling — and intent modeling is where most recommendation systems are weak.
Intent as a first-class signal
Users arrive at different moments with different needs. A user exploring a category has different intent than a user with a specific item in mind. A returning user has different needs than a first-session user. Recommendations that don't model intent serve the same content to all users regardless of context — which means they're relevant for some users some of the time and irrelevant for most users most of the time. The product work is segmenting these intent modes and serving each one differently.
The context problem in collaborative filtering
Collaborative filtering is powerful but context-blind. It can tell you what users similar to this user have bought, but not what this user is likely to buy right now, given their current session context, time of day, recent browsing, and entry point. Adding context as a feature to a collaborative filtering model, or using a contextual bandits approach, is often the single highest-impact intervention available to recommendation teams.
Product moment design
Where you place a recommendation is as important as what you recommend. Recommendations placed in a session where the user is actively evaluating convert better than recommendations placed at moments of completion. The product discipline is mapping the user journey, identifying the high-intent moments, and designing recommendation surfaces specifically for those moments — rather than adding recommendations everywhere and hoping for the best.
Revenue-oriented recommendation systems are built by teams that refuse to accept engagement as a proxy for business value. The technical work — intent modeling, context features, product moment design — is tractable. The organizational work — convincing stakeholders to optimize for the right metric — is often harder.