Product recommendations drive an estimated 35% of Amazon's revenue. For eCommerce businesses of any size, intelligent recommendations increase average order value, improve product discovery, and create a more engaging shopping experience. The technology behind effective recommendations has become increasingly accessible, and you do not need Amazon's resources to implement it. Here is how product recommendation engines work and how to deploy them effectively.
How Recommendation Engines Work
There are three fundamental approaches to product recommendations, each with different strengths:
- Collaborative filtering: Recommends products based on what similar users have purchased or viewed. "Customers who bought X also bought Y." This works well with large datasets but struggles with new products (cold start problem) and new users with no purchase history.
- Content-based filtering: Recommends products similar to what the user has already viewed or purchased, based on product attributes like category, brand, colour, or price range. "Because you viewed X, you might like Y." This handles new products well but tends to produce narrow, repetitive recommendations.
- Hybrid systems: Combine both approaches to compensate for each method's weaknesses. Most production recommendation systems use hybrid models that blend collaborative and content-based signals with contextual data (time of day, season, device type) for the most relevant results.
Recommendation Placement Strategies
Where you place recommendations matters as much as their quality. Product pages should show "Frequently bought together" (cross-sell) and "Customers also viewed" (discovery). Cart pages should display complementary items that add value to the current purchase — not competing alternatives. Homepage recommendations should reflect the returning visitor's browsing history and preferences. Post-purchase emails should suggest accessories, refills, or complementary products based on what was just purchased.
Each placement serves a different goal. Cross-sell recommendations on product and cart pages increase average order value. "You may also like" sections on product pages reduce bounce rate by offering alternatives. Homepage personalisation increases engagement and return visit frequency. Tailor your recommendation algorithm and presentation to match the goal of each placement.
Implementation Options
For most eCommerce businesses, building a recommendation engine from scratch is unnecessary. Third-party services like Nosto, Barilliance, and Clerk.io provide plug-and-play recommendation widgets that integrate with Shopify, WooCommerce, and Magento. These services handle data collection, model training, and serving recommendations via JavaScript widgets or API endpoints.
For businesses with custom platforms or specific requirements, cloud ML services from AWS (Amazon Personalize), Google Cloud (Recommendations AI), or open-source frameworks like LensKit provide the building blocks. Amazon Personalize is particularly interesting — it uses the same technology that powers Amazon's own recommendations and is accessible via API without requiring ML expertise.
Data Requirements
Recommendation engines need data to be effective. At minimum, you need product catalogue data (attributes, categories, prices) and user interaction data (views, add-to-carts, purchases). The more data points you collect, the better recommendations become. Session behaviour (click paths, dwell time), search queries, wishlist additions, and review interactions all enrich the recommendation model.
New stores face the cold start problem — insufficient data to make meaningful recommendations. Handle this with rule-based fallbacks: show bestsellers to new visitors, trending products by category, or manually curated collections. As your data grows, gradually shift from rule-based to ML-driven recommendations.
Measuring Recommendation Performance
Track click-through rate on recommendation widgets, the percentage of revenue attributable to recommended products, and the impact on average order value. Run A/B tests comparing different recommendation algorithms, placements, and widget designs. The benchmark to aim for is 10-30% of total revenue originating from product recommendations, depending on your catalogue size and industry.
At Born Digital, we implement recommendation systems as part of broader eCommerce personalisation strategies. The most effective implementations combine algorithmic recommendations with merchandising rules — ensuring that recommended products align with business priorities like clearing seasonal inventory, promoting high-margin items, or supporting new product launches. The technology is the enabler, but the strategy determines the business impact.