Product Recommendations in E-commerce: How They Work
Advanced4 minInnovialabs editorial team
In short
Recommendations grow from simple rules to models built on behavior and embeddings; they belong on product pages, carts and emails, and their value is proven only by an A/B test on revenue per visitor.
How product recommendations work
A recommender answers one question: "what else should this person see right now?" There are four main approaches, and real stores usually combine them:
| Approach | Signal source | Strength | Weakness |
|---|---|---|---|
| Rules | Merchandiser decisions | Control, fast start | Doesn't scale to a large catalog |
| Collaborative filtering | Other people's purchases and views | Finds non-obvious links | Needs history, cold-start problem |
| Content-based | Product attributes: category, brand, price | Works for new products | Suggests items that are too similar |
| Embeddings | Vectors for products and users | Flexible, semantic similarity | Harder to build and maintain |
Start with rules and simple statistics, and add models once you have data and a way to measure impact.
Rules and cross-sells
The fastest start is manual or semi-automatic rules:
- accessories: a case for a phone, a filter for a coffee machine;
- "frequently bought together" based on co-purchases in orders;
- free-shipping fillers — inexpensive items in the cart;
- exclusions: never recommend out-of-stock items or what the customer already bought.
Rules are easy for the business to understand and explain, but hard to maintain across tens of thousands of products.
Collaborative filtering
The idea: if people who bought A often bought B, show B to buyers of A.
- Item-to-item — computes item similarity from shared interactions. Stable and well suited to product pages.
- Matrix factorization — decomposes the user × item matrix into latent factors and predicts interest in unseen items.
Weight signals differently: a purchase counts more than add-to-cart, which counts more than a view. Account for popularity: without normalization every recommendation collapses into best-sellers.
Content-based and embedding approaches
Content-based methods compare products by attributes: category, brand, material, price tier. Embeddings go further — they turn titles, descriptions and images into vectors so that semantically similar products end up close together. Vector indexes handle the nearest-neighbor search.
Behavioral embeddings can be built from sessions too: products often viewed one after another get similar vectors.
Handling cold start
- New product — use content and embedding recommendations until behavior accumulates.
- New user — show category best-sellers, trends and items based on the current session.
- New store — merchandiser rules and best-sellers while data is scarce.
A good practice is a hybrid: a behavioral model that falls back to content-based results when data is thin.
Where to place them
| Placement | What to show |
|---|---|
| Product page | Similar items and "frequently bought together" |
| Cart | Add-ons, accessories, low-priced items |
| Home and catalog | Personal picks, continue browsing |
| Email and push | Abandoned cart, consumable refills, items similar to past purchases |
| Empty search results | Popular items in related categories |
Don't overload the page: several blocks with different logic compete for attention and make each one harder to evaluate.
Measuring impact
Block CTR says nothing about money: clicks may simply move from elsewhere on the site. You need an A/B test:
- Split users (not sessions) into control and variant.
- Primary metric — revenue per visitor; secondary — conversion, average order value, share of items from recommendations.
- Decide test duration and sample size in advance.
- Watch guardrail metrics: returns and page speed.
Test channels separately: a recommendation in an email and on a product page are different experiments.
Common mistakes
- Recommending out-of-stock items.
- Showing what the customer already bought, unless it's a consumable.
- Judging a model only by offline metrics, without an online test.
- Never retraining — seasonality makes a model stale quickly.