Building AI-Driven Personalisation Engines for Retail
Customers who receive personalised experiences spend more, return more often, and are more likely to recommend the brand. Yet despite years of investment in recommendation engines, many retailers find that their personalisation capabilities plateau — delivering obvious suggestions rather than genuinely useful ones. The difference between good and great personalisation lies in data strategy, model design, and real-time execution.
Beyond Collaborative Filtering
Classic collaborative filtering — recommending products based on what similar users bought — remains useful but has well-known limitations: it struggles with new users, new products, and seasonal shifts. Modern personalisation systems combine collaborative filtering with content-based signals, real-time session behaviour, and contextual features like time of day, device type, and current promotions.
Real-Time vs. Batch Personalisation
Batch-computed recommendations updated nightly are no longer sufficient for competitive retail experiences. Real-time personalisation that responds to the current session — adapting as a user browses, searches, and adds to cart — significantly outperforms static recommendations. Building the streaming data infrastructure to support real-time inference adds engineering complexity but delivers measurable uplift.
The Cold Start Problem and How to Solve It
New users and new products have no interaction history. Solving cold start requires a combination of onboarding flows that capture explicit preferences, heuristic-based fallbacks, and content-based models that can make recommendations from product attributes alone. For new products, popularity signals from similar items and editorial curation provide useful starting points.
Measuring What Actually Matters
Click-through rate is an easy metric to optimise but can lead personalisation systems toward clickbait-style recommendations that frustrate customers. Revenue per session, basket size, and long-term retention are more meaningful business metrics. A/B testing infrastructure that can measure these outcomes correctly — accounting for novelty effects and selection bias — is as important as the recommendation model itself.
Other Stories.
See all cases
We'd love
to help.
