Recommendation engines that actually lift revenue
Most recommendation widgets recommend what the customer already bought. Real personalisation predicts the next thing they want before they know it — and does it fast enough to matter while they're still on the page.
The difference between a recommender that lifts revenue and one that's just decoration comes down to signal, speed, and measurement. Get those three right and personalisation becomes one of the highest-ROI systems in the whole business.
Use behaviour, not just purchases
Clicks, dwell time, cart adds, and searches carry far more intent than the sparse signal of completed purchases. Blend real-time session behaviour with long-term preferences so the model reacts to what the customer is doing right now, not what they did last quarter.
Serve in real time or don't bother
- Precompute embeddings offline, but rank candidates in real time against the live session.
- Cap end-to-end latency to a few dozen milliseconds — a recommendation that lands after the page does nothing.
- Fall back gracefully to popular-in-category when you have no signal, so new visitors still see something relevant.
Measure incremental lift, not clicks
A recommender can show high click-through while adding zero revenue, because it recommends what people would have bought anyway. Always test against a holdout and measure incremental lift. If the model can't beat 'no recommendations' on real money, it isn't earning its place.
The best recommendation isn't the one that gets clicked. It's the one that wouldn't have happened without it.
Respect the customer's limits
Over-personalisation feels like surveillance. Diversify results, avoid creepy inferences, and give customers control. Trust compounds — and a customer who trusts your recommendations buys from them for years.
Want this built for your team?
We ship production-grade AI like this across every industry — in weeks, not months.