Product recommendations for ecommerce
How they work, where to place them, and what to look for in a recommendation engine
Product recommendations are automated suggestions that help shoppers discover relevant products based on their behavior, the items they're viewing, and broader purchase patterns. The most cited reference point is McKinsey's finding (2017, citing 2013 data) that 35 percent of what consumers purchase on Amazon comes from product recommendations. They work because they solve a fundamental problem: shoppers can't browse your entire catalog, so the system surfaces the most relevant products for each individual at each moment.
This guide covers the types of recommendations, where to place them, how the algorithms work, and how to evaluate a recommendation engine for your store.
Types of product recommendations
Similar products
Alternatives to the product being viewed, same category, similar attributes, comparable price range. Helps comparison shoppers find the right option. "If you like this, you might also consider..."
Frequently bought together
Products commonly purchased in the same order. A phone case with a screen protector. Running shoes with performance socks. This is the primary cross-sell driver and the recommendation type with the most direct AOV impact.
Trending and popular
Products with rising purchase velocity or view counts. Useful for new visitors where there's no behavioral data to personalize with. Also effective for FOMO-driven categories like fashion and limited editions.
Personalized for you
Recommendations based on the individual shopper's browsing history, purchase patterns, and real-time behavior. The most effective type for returning visitors. Requires behavioral data from at least one session to begin working.
Recently viewed
A simple but effective pattern, showing products the shopper has already looked at. Reduces friction for shoppers who are comparing options across multiple sessions. Low algorithmic complexity but high utility.
Contextual recommendations
Recommendations that change based on context, time of year (winter coats in November), day of week (party supplies on Fridays), or external events (sun cream when the weather forecast is hot). Requires integration with contextual data sources beyond purchase history.
Where to place recommendations (and why)
| Placement | Best recommendation types | Primary impact |
|---|---|---|
| Product detail page | Similar, frequently bought together | Lifts AOV by adding complementary items to consideration |
| Cart page | Complementary, add-ons | Increases items per order at the highest-intent moment |
| Homepage | Personalized, trending | Engagement, return visits, faster path to relevant categories |
| Category page | Popular in category, personalized | Product discovery, surface depth beyond top sellers |
| Personalized, replenishment | Drives repeat purchase and reactivation | |
| Zero-result search | Popular, recently viewed | Recovers sessions that would otherwise exit empty-handed |
CTR ranges by placement vary widely with design, product mix, and traffic source, set your own baseline from your first month and track lift over time, rather than chasing industry-wide percentages.
How recommendation algorithms work
Collaborative filtering
"People who bought X also bought Y." Analyzes purchase patterns across all shoppers to find correlations. Strengths: discovers non-obvious relationships (beer + diapers). Weaknesses: cold-start problem (new products with no purchase history get no recommendations), popularity bias (tends to recommend bestsellers).
Content-based filtering
"Products with similar attributes to what you've viewed." Uses catalog data, categories, brands, colors, price ranges, materials, to find similar items. Strengths: works for new products immediately, no cold-start problem. Weaknesses: limited to explicit attributes, can't discover the unexpected connections that collaborative filtering finds.
Hybrid and deep learning approaches
Modern systems combine collaborative and content-based signals with behavioral data (browsing patterns, time on page, scroll depth) and contextual data (time of day, device, location). Deep learning models encode all these signals into vector representations, enabling nuanced matching that captures both explicit attributes and implicit behavioral patterns.
Platform-specific considerations
Shopify
Shopify's built-in recommendations are basic, rule-based, limited to "related products" in the same collection. Third-party apps integrate via the Storefront API. Look for apps that use real behavioral data, not just product metadata. Check whether the app works with your theme and doesn't slow down page load.
Magento / Adobe Commerce
Adobe Commerce includes Product Recommendations powered by Adobe Sensei. It's capable but tightly coupled to the Adobe ecosystem. Third-party options offer more flexibility and often better relevance. Implementation via widgets or API, check compatibility with your theme and custom modules.
WooCommerce
WooCommerce has minimal built-in recommendations (related products by category/tag). The plugin ecosystem offers many options, but quality varies widely. Prioritize plugins that use behavioral data over simple category matching, and check performance impact, some plugins add significant page weight.
Headless / custom
API-first recommendation engines (via REST or GraphQL) give full control over placement, design, and logic. You build the frontend; the engine provides the intelligence. This approach requires more development effort but offers the most flexibility and best performance (no third-party widget overhead).
How to evaluate a recommendation engine
- 1. Relevance. Do recommendations make sense for your catalog? Test with your actual products. "Frequently bought together" should show genuine complements, not random items from the same category.
- 2. Cold-start handling. What happens for new products with no purchase history? What about new visitors with no behavioral data? The engine should have fallback strategies that still deliver relevant results.
- 3. Merchandising controls. Can you boost, block, or pin specific products in recommendation slots? Can you exclude out-of-stock items, low-margin products, or specific brands?
- 4. Placement flexibility. Can you place recommendations anywhere on your site, or only in pre-defined locations? API-based engines offer the most flexibility; widget-based engines are faster to implement but more constrained.
- 5. Cross-channel capability. Can the same engine power on-site recommendations AND email recommendations? Sharing behavioral data across channels produces better results than using separate systems.
Frequently asked questions
What are product recommendations in ecommerce?
Product recommendations are automated suggestions shown to shoppers based on their behavior, the product they're viewing, or broader trends. They appear throughout the shopping journey, homepage, category pages, product detail pages, cart, checkout, and email. The widely cited McKinsey benchmark (2017, citing 2013 data) is that 35 percent of what consumers purchase on Amazon comes from recommendations, a long-standing reference point for what mature, large-scale recommendation systems can drive.
What types of product recommendations are there?
The main types are: similar products (alternatives to what you're viewing), frequently bought together (complements), trending/popular items, personalized for you (based on individual behavior), recently viewed, and new arrivals in categories you browse. Each type serves a different intent, similar products help comparison shoppers, while frequently bought together drives cross-sell.
Where should I place product recommendations on my store?
The highest-converting placements are: product detail page (similar and complementary items), cart page (add-ons and accessories), homepage (personalized picks for returning visitors, trending items for new visitors), category pages (contextual suggestions), post-purchase email (cross-sell and replenishment), and search results page (for zero-result queries). Start with PDP and cart, they have the highest conversion impact per impression.
How do recommendation algorithms work?
Three main approaches: collaborative filtering (people who bought X also bought Y, uses purchase patterns across all shoppers), content-based filtering (products with similar attributes to what you've viewed, uses catalog data), and hybrid approaches that combine both with behavioral signals. Modern systems use vector embeddings to represent products in high-dimensional space, enabling nuanced similarity matching beyond simple rules.
What is the difference between rule-based and AI recommendations?
Rule-based recommendations follow explicit logic: 'show bestsellers', 'show products from the same category', 'show items in the same price range.' AI recommendations learn from data: they discover that customers who buy hiking boots tend to buy merino wool socks 2 weeks later, even though those products are in different categories and price ranges. Rule-based is predictable but limited; AI discovers patterns humans miss but requires sufficient data to learn from.
What conversion rates should I expect from product recommendations?
Recommendation CTR varies widely by placement and relevance, sponsored product slots in search results convert differently from sidebar carousels on a product detail page. Conversion rates from recommendation clicks are usually higher than site average because the shopper has already shown intent by engaging with a suggestion. Measure direct recommendation revenue and assisted revenue (sessions that touched a recommendation before converting) separately, the assisted figure is usually larger and easier to under-credit.
See recommendations powered by Product Intelligence
Hello Retail's recommendation engine uses Product Intelligence to deliver personalized suggestions across web, email, and search.
See Hello Retail recommendations