Personalization at Scale: Recommendation Engines That Actually Lift Revenue

A basic collaborative-filtering "customers also bought" widget is easy to bolt onto any store and it plateaus fast, because it optimizes for popularity, not for the specific customer looking at the screen. The recommendation systems that measurably lift average order value go further: they combine several signals and, critically, they are tuned against a real business metric, not just click-through rate on the widget itself.

Signals worth combining

• Collaborative filtering — what similar customers bought together, the baseline everyone starts with.
• Content-based similarity — attributes of the item itself (category, price band, material), useful for new products with no purchase history yet.
• Session behaviour — what this specific visitor has viewed and lingered on in the last few minutes, which often outweighs their entire purchase history for intent.
• Inventory and margin awareness — a recommendation engine blind to stock levels and margin will happily promote a product you are out of, or one that quietly erodes your margin every time it converts.

That last point is the one pure ML-focused teams miss most often. A recommendation is not just a relevance problem, it is a business decision, and the engine should be able to weight toward inventory you need to move and margin you want to protect, not just statistical similarity.

Optimize for AOV lift, not click-through rate

A widget with a high click-through rate that recommends items customers were going to buy anyway adds nothing — it is measuring engagement, not incremental revenue. We run recommendation placements as controlled experiments against a genuine incremental-revenue metric (holdout groups, not just before/after), because it is common for a "successful" widget by click-through to show zero measurable AOV lift once isolated properly.