Almost every recommendation engine is optimizing for the wrong thing. Click-through rate is easy to measure and easy to game. Showing a customer a product they're very likely to click on — something they almost certainly already know about — creates a click, adds it to the logged data, and contributes nothing to revenue. The customer leaves without buying, or buys the one thing they came for, and never returns. The recommendation engine reports a high engagement rate and calls it success.
The metric that matters is lifetime value — what a customer spends over the entire relationship with the brand, not what they spent in this session. A recommendation engine that lifts LTV often looks less impressive on a click-rate dashboard. It shows customers things they didn't know they wanted but genuinely need, which converts less immediately but far more durably.
The three types that actually move revenue
Complementary recommendations show items that work better when used together. A camera body and a lens. A notebook and a pen refill. A skincare cleanser and a matching moisturiser. These are the highest-converting recommendations in any category because they address a real, active need — the customer is already in buying mode, and you're making the purchase more complete. Average order value goes up; return rates go down because customers have what they actually need.
Replenishment recommendations are the sleeper metric in eCommerce. If you know a customer bought a 90-day supply of a supplement 85 days ago, showing them a reorder recommendation at day 82 is one of the highest-ROI personalizations you can run. The timing is the intelligence. Replenishment is what turns a customer who bought once into a customer who buys every quarter — that's the compounding value that LTV captures and click-rate misses entirely.
Discovery recommendations are the hardest to get right and the most valuable when you do. Showing a customer something genuinely new — something outside their purchase history but aligned with an emerging need — is what builds loyalty. It's the digital equivalent of a great shop assistant who remembers what you bought last time and says "actually, you might want to look at this." When it works, customers feel understood. When it's wrong, it feels intrusive. The difference is data quality and model sophistication.
The recommendation engine that lifts LTV often looks less impressive on a click-rate dashboard. Build for the right metric from the start.
The cold-start problem
Every recommendation system suffers from the cold-start problem: new customers have no purchase history, so there's nothing to personalize against. The temptation is to show bestsellers — it's safe, it converts reasonably well, and it requires no data. The problem is that bestsellers recommendations actively destroy the data you need to personalize later. Every new customer shown a bestseller gets the same first experience, and the resulting signal is useless for understanding individual preference.
A smarter cold-start approach uses onboarding signals — stated preferences, browsing behaviour, session context — to seed a lightweight initial model before the first purchase. Even one or two signals narrow the recommendation space meaningfully. The goal isn't perfection on visit one; it's collecting enough signal to personalize by visit two.
What to measure
Track repeat purchase rate, average order value for recommendation-influenced sessions, and 90-day LTV cohorted by acquisition channel. If your recommendation engine is working, you'll see repeat purchase rate climb — customers coming back more often because previous recommendations felt accurate. Average order value climbs because complementary recommendations are being accepted. And 90-day LTV diverges between customers who engaged with recommendations and those who didn't.
If those three metrics aren't moving, the click-rate number is a vanity metric and the engine needs to be rebuilt around a different objective function. That's the conversation we start with before writing a single line of recommendation logic.
