"Customers who bought this also bought" is the baseline. Modern AI personalisation engines read intent signals in real time and surface the right product at the right moment — lifting revenue per visitor by 15–30%.
Product recommendation engines have existed since Amazon pioneered collaborative filtering in the 1990s. What's changed is the signal richness available today — real-time browsing behaviour, search queries, cart abandonment patterns, purchase history, price sensitivity signals, and external context like time of day and device type — and the model sophistication to act on all of it simultaneously.
The Signals That Drive Personalisation Quality
- Explicit signals — past purchases, saved wishlists, explicitly stated preferences
- Implicit browse signals — products viewed, time spent on each page, scroll depth, zoom events
- Search intent — the specific words a customer uses in site search reveal intent more clearly than browse behaviour
- Abandonment patterns — what a customer added to cart but didn't buy reveals price sensitivity and comparison behaviour
- Session context — time of day, device, referral source, and campaign context all inform recommendation strategy
- Segment signals — cohort-level patterns from similar customers fill in the picture for new or low-activity users
Where Personalisation Lifts Revenue Most
The highest-impact personalisation touchpoints in a typical eCommerce journey are the product detail page (complementary and "complete the look" recommendations), the cart page (last-chance add-ons), the post-purchase email (replenishment and cross-category recommendations timed to purchase frequency), and the home page for returning visitors (curated based on browse history rather than generic bestsellers).
Across eCommerce implementations we've built, personalised product recommendations consistently account for 15–30% of total revenue — with the highest performers being cart-page and post-purchase email recommendations where purchase intent is already high.
Cold Start: Handling New Users
A new visitor with no history is the hardest case for personalisation. The practical solution is a warm-up sequence: start with category-level bestsellers, adapt based on the first 2–3 pages browsed in the session, and use any available context (referral source, UTM campaign, geolocation) to make initial recommendations less generic while individual behaviour history builds.
