Ecommerce Product Recommendations: How Recommendation UX Should Work
How ecommerce product recommendations should work: related, complementary, alternatives, recently viewed and personalized, placement, relevance and explainability.
Quick answer
Product recommendations help shoppers discover products they wouldn't otherwise find and complete their purchase. Match the type to the moment: similar alternatives while shoppers are choosing, complementary items near add-to-cart and in the cart, bestsellers and recently viewed items where there's little context, and replenishment after purchase. Combine algorithms with merchandising rules, exclude unavailable or unsuitable products, explain each module with a clear title, keep modules fast and accessible, and measure impact against a holdout rather than trusting attributed clicks.
Recommendation Types and Their Jobs
| Type | Shopper's situation | Typical placement |
|---|---|---|
| Similar / alternatives | Still choosing; this item isn't quite right | Product page, below the buy area; empty search |
| Complementary | Decided; needs things that go with it | Near add-to-cart, cart |
| Frequently bought together | Building a set or kit | Product page, cart |
| Bestsellers / trending | No context yet | Homepage, category, empty states |
| Recently viewed | Returning or comparing | Homepage, product page, search |
| Personalized picks | Returning with history | Homepage, email |
| Replenishment | Consumable running out | Email, account, homepage |
How Recommendations Are Generated
| Approach | How it works | Strengths | Limits |
|---|---|---|---|
| Merchandiser rules | Manually chosen or rule-based lists | Control, brand intent | Doesn't scale; goes stale |
| Co-purchase / co-view | Items bought or viewed in the same session or order | Captures real complements | Needs volume; new items excluded |
| Attribute similarity | Items with similar category, attributes, price | Works for new products | Depends on clean product data |
| Collaborative filtering | Patterns across many shoppers' behavior | Finds non-obvious relationships | Cold start; needs lots of data |
| Hybrid | Combination with business rules | Balanced | More complex to tune |
Pro tip
Most stores get better results from a simple algorithm with good product data and sensible exclusions than from a sophisticated one with messy data.
The Cold Start Problem
New products and first-time visitors have no behavioral history. Fill the gap with attribute-based similarity, category bestsellers and merchandiser picks, then let behavioral signals take over as data accumulates. Structured product attributes are what make this work. See ecommerce product discovery.
Placements Across the Journey
The diagram above maps recommendation types to pages. The principle: recommend alternatives to shoppers who are deciding and complements to shoppers who have decided. On the product page, place them where they support the decision rather than compete with the buy area; see product page design. Putting upsells between a shopper and the checkout button often costs more than it adds.
Designing Recommendation Modules
- Titles that say why: “Goes well with”, “Similar styles”, “Customers also bought”
- Product image, name, price and rating visible on each card
- One row or a short carousel; don't stack several modules of the same type
- Quick add for low-consideration complements, with variant selection where needed
- Carousels usable by touch, keyboard and screen readers
- Loaded without delaying the main content or shifting the layout
- Rendered as real links so they also support internal linking
Recommendations getting clicks but not sales?
ZSpace reviews recommendation types, placements and measurement, and redesigns modules around shopper intent.
Merchandising Controls
Algorithms need guardrails. Exclude out-of-stock products, products with high return rates and items that don't make sense together (a second sofa as a “complement” to a sofa). Consider margin, especially for complements. Pin strategic items sparingly and review rules regularly.
Measuring Honestly
Attributed revenue, meaning revenue from orders containing a recommended item that was clicked, overstates impact because many of those shoppers would have bought anyway. Keep a holdout group that sees no recommendations (or a simple baseline), and compare revenue per session, average order value, conversion and returns between groups.
| Metric | Use |
|---|---|
| Module click-through | Is the module relevant and noticed? |
| Add-to-cart from module | Does it influence choice? |
| Revenue per session vs holdout | Is there incremental value? |
| AOV and items per order vs holdout | Is it building baskets? |
| Returns on recommended items | Are recommendations setting the right expectations? |
Testing Recommendation Strategies
Test one variable at a time: type (similar vs complementary), placement (above or below reviews or in the cart), number of items or algorithm. Use revenue per session as the primary metric and conversion as a guardrail, since a module can raise order value while distracting some shoppers from buying. See ecommerce A/B testing.
Platform Notes
Many platforms include recommendations. On Shopify, the platform generates product recommendations and the Search & Discovery app lets merchants customize related and complementary products; Shopify notes complementary products must be active and in stock to appear (Shopify Help Center). Third-party engines add personalization and more control for larger catalogs. See Shopify product recommendations.
Want recommendations that help shoppers find more?
Talk to ZSpace about CRO, module design and recommendation systems.
Recommendation Architecture
Most recommendation systems follow the same pipeline, whether built in-house or provided by an app. Understanding it helps diagnose poor recommendations: most problems come from data, filtering or placement rather than the algorithm.
signals : orders, views, carts, searches, product attributes
candidates: co-purchase, similar attributes, same category, trending, recently viewed
filter : in stock, available in market, not already bought/in cart, merchandising rules
rank : relevance to context (page, customer), diversity, margin rules
place : module on page with a clear label and reason
measure : clicks, add to cart, revenue vs holdoutExplainability: Tell Shoppers Why
Recommendations are more useful when shoppers understand why they're shown: “Goes with your jacket”, “Similar, lower price”, “Because you viewed trail shoes”, “Customers who bought this also bought”. Labels help shoppers judge relevance and reduce the sense of being watched. Avoid vague “You may also like” where a specific reason is available, and never imply a reason that isn't true.
| Type | Label example |
|---|---|
| Complementary | Complete the look / Works with this |
| Alternatives | Similar styles / Compare with |
| Recently viewed | Recently viewed |
| Personalized | Picked for you, based on your orders |
| Popular | Bestsellers in running |
Worked Example: Fixing Product Page Recommendations
An illustrative scenario: an outdoor gear store's product pages show one “You may also like” carousel above the reviews, filled mostly with near-identical items and some out-of-stock products. The team moves recommendations below product details, splits them into “Works with this” (accessories from co-purchase data, filtered for stock) and “Compare similar” (same category, different price points), adds labels, and tests the change against a holdout. They measure add-to-cart from modules and product page conversion, watching that the modules don't distract from the main product. See AI product recommendations and cross-selling.
Common Mistakes
- Several near-identical modules on one page
- Upsells between the shopper and checkout
- Recommending out-of-stock or just-purchased items
- Vague titles like “You may also like” everywhere
- Judging success by attributed revenue alone
- Heavy scripts that slow the product page
Conclusion
Recommendations increase discovery when the type fits the moment, the data is clean, merchandising guardrails are in place and impact is measured against a holdout. For the wider picture of relevance, see ecommerce personalization.
Common questions
Modules that suggest products to a shopper, such as similar items, complementary items, bestsellers or personalized picks, to help them find something to buy or complete a purchase.