Ecommerce Product Recommendations: Types, Placement and Measurement

Ecommerce product recommendations are product lists chosen automatically for a shopper, such as "customers also bought", "similar items" or "recommended for you". They work when each type is placed where it matches the shopper's goal on that page: alternatives on a product page, complements in the cart, discovery on the home page. Measure them with controlled tests and revenue per visitor, not with the revenue that happens to pass through the widget.

This guide explains the main recommendation types, gives a placement grid showing which type belongs on which page, and sets out how to measure the result honestly.

The main recommendation types

Most engines, including Adobe's Product Recommendations service for Adobe Commerce, group recommendation types by the data they use. Adobe's documentation, for example, lists personalized types (Recommended for you, Recently viewed), cross-sell and up-sell types (Viewed this, viewed that; Viewed this, bought that; Bought this, bought that; More like this; Visual similarity), popularity types (Most viewed, Most purchased, Most added to cart, Trending) and high-performing types based on view-to-purchase and view-to-cart conversion. Other engines use different names for similar ideas.

Behavioral (collaborative filtering)

"Bought this, bought that" and "viewed this, viewed that" learn from aggregate shopper behavior. They surface real-world relationships that nobody wrote down, such as which accessories go with a camera. They need enough traffic and purchase history to be reliable, and they struggle with new products.

Content-based (similarity)

"More like this" uses product attributes such as category, brand, description and specifications. Visual similarity uses images. These work from day one for new items because they do not depend on behavior, but they only find products that resemble each other, not products that complement each other.

Popularity and trending

Best sellers, most viewed and trending lists are simple and robust. They are useful fallbacks when there is no personal history, and on landing pages for new visitors. Adobe notes that its popularity types use the last seven days of storefront behavior.

Personalized

"Recommended for you" blends a shopper's own browsing and purchase history with patterns from similar shoppers. It is the most powerful type for returning customers and the least useful for first-time anonymous visitors.

Rule-based and merchandised

Manually curated bundles, "complete the look" sets and compatibility rules (this filter fits that model) are still valuable. Algorithms are weak at hard constraints, and a wrong compatibility suggestion is worse than none.

Placement grid: recommendation type by page

The grid below matches each type to the pages where it usually fits the shopper's intent. "Primary" means a strong default; "Secondary" means worth testing; a dash means it usually works against the page's goal.

Three placements deserve extra attention:

• Out-of-stock product pages. Similar items should appear high on the page, because the shopper cannot buy what they came for.
• Cart and checkout. Keep it to a few relevant, lower-priced add-ons. Anything that sends the shopper away from checkout should be tested carefully.
• Zero-results search pages. Trending or category best sellers give the shopper somewhere to go instead of a dead end.

The order confirmation page and post-purchase emails are also good places for "bought this, bought that" and replenishment suggestions, since the purchase decision is already made.

Data and setup requirements

• Clean catalog attributes. Similarity types depend on them; inconsistent categories and missing attributes produce strange suggestions.
• Reliable behavioral events. Product views, add-to-cart and purchases must be captured correctly, including on headless storefronts, where Adobe notes that eventing must be set up separately.
• Enough history. Behavioral types need days to weeks of activity before they are useful; Adobe's documentation notes that new or low-traffic stores may see limited results and that engines fall back to popularity types for shoppers without history.
• Exclusion rules. Filter out of-stock items, products already in the cart, items already purchased (for durable goods) and anything you cannot ship to the shopper's region.
• Consent. Personalized recommendations rely on behavioral data. Make sure tracking respects the shopper's consent choices in your markets.

How to measure recommendations

The most common mistake is reporting "revenue from recommendations", meaning the value of orders that contained a clicked recommended item. That number counts sales that would often have happened anyway, so it makes every widget look successful.

Better practice:

1. Run A/B tests per placement. Compare a page with the recommendation unit against the same page without it, or against a simpler type such as best sellers.
2. Use revenue per visitor as the primary metric. It captures both conversion and order value, and it catches cases where a widget distracts shoppers from buying.
3. Watch secondary metrics. Click-through rate on the unit, average order value, items per order, and for cart placements, checkout completion rate.
4. Segment the results. New and returning visitors, mobile and desktop, and major categories often behave differently.
5. Check margin, not just revenue. Recommending heavily discounted items can raise revenue and lower profit.

Run tests long enough to cover full weekly cycles and promotion periods. Our ecommerce conversion rate optimization guide covers test design in more detail.

Recommendations as part of personalization

Recommendations are usually the first personalization most stores deploy, because they are self-contained and easy to test. The next steps, such as personalized search ranking, content and email, need more data and more governance. See our guide to ecommerce personalization for a maturity ladder, and AI in ecommerce for how recommendations fit with search, chat and content generation.

FAQ

How many recommendation units should a page have?

Usually one or two per page, each with a distinct purpose. Several units with overlapping products dilute attention and slow the page.

Do recommendations slow down my store?

They can if widgets load synchronously or render large image carousels above the fold. Load them asynchronously, reserve their space to avoid layout shift, and keep them below the main content where possible.

Can I use recommendations on Magento Open Source?

Adobe positions its Product Recommendations service for Adobe Commerce merchants; check Adobe's current eligibility terms for your license. Magento Open Source stores typically use built-in related products, up-sells and cross-sells, a third-party extension, or an external recommendation service.

Next steps

If you want help choosing a recommendation engine, setting up clean events, or designing tests for each placement, Mistri's AI for ecommerce services cover planning, integration and measurement.

Sources

Adobe Experience League: Product Recommendations types; Adobe Experience League: Product Recommendations overview.