Product recommendations are blocks that suggest other items to a shopper based on what they are viewing, what is in their cart, or what similar shoppers bought.
In our database they win most often in the cart and on sold-out product pages, where they add to the order or rescue a dead end.
They lose when they crowd the checkout button or push the main product's buy box down the page.
This page covers how recommendation engines pick products and where recommended products have won and lost in real A/B tests.
It sits inside our personalization section, and overlaps with upsells: recommendations are the engine, and upsells and cross-sells are two of the things you can do with them.
How a product recommendation engine works
A product recommendation engine is the software that decides which products to show.
Most engines use one of three approaches, or a blend of them.
| Approach | What it uses | Good at | Weak at |
|---|---|---|---|
| Item-to-item collaborative filtering | Which products customers buy together | "Frequently bought together" and cart add-ons | New products with no sales history |
| Content-based (similar attributes) | Product descriptions, tags, categories | "Similar products" when an item is sold out | Suggesting accessories that look nothing like the main item |
| Rules and manual picks | A merchant's own choices | Small catalogs, required accessories, margin control | Scaling to thousands of products |
The best-known version of the first approach is Amazon's.
In a 2003 IEEE Internet Computing paper, Greg Linden, Brent Smith and Jeremy York described item-to-item collaborative filtering: rather than matching a shopper to similar customers, it matches each item the shopper bought or rated to similar items, using a similar-items table built from products customers tend to purchase together.
The paper's key point was speed. Amazon built the expensive similar-items table offline, so the live lookup depended only on how many items the shopper had bought, not on the size of the catalog or customer base.
Related vs complementary recommendations
The most useful distinction for a store is between related products (alternatives to the item) and complementary products (things that go with it).
Shopify builds this split into its platform. Its Product Recommendations API accepts two intents, related and complementary, and returns up to 10 products.
According to Shopify's help docs, related products are generated automatically from purchase history, product descriptions (English storefronts only) and related collections as a fallback.
Complementary products are picked by hand in the Search & Discovery app, up to 10 per product.
That matters for testing. A related block on a product page offers the shopper a way out of the product they are on.
A complementary block offers a way to add to it.
The two can produce opposite results in the same spot.
Does the engine itself matter?
Sometimes, yes. Blend Commerce ran a test for Stone Creek Coffee that kept the design and position identical and swapped only the source of the ten recommendations, from Shopify's native API to Rebuy's engine.
The Rebuy version showed +23% conversion rate, +25% revenue per visitor and +84% subscription revenue per visitor.
Treat that with care. The test ran 4 days on 3,452 visitors and stopped at an 89.5% probability to beat control, so it is a promising signal rather than proof.
A sample size calculation before you start will tell you how long a test like this really needs.
Vendor-published engine swaps point the same way but are harder to trust.
Visually reported +11.7% conversion rate for Clare V. after swapping cart recommendations from another tool to its own, with no sample size or significance stated.
AB Tasty reported +21% homepage conversion for Bricomarché after replacing manual picks with AI recommendations, but that was a before-and-after rollout, not a controlled test.
The honest summary: a better engine can help, but placement and page layout show bigger, more consistent effects in our data.
For more on algorithmic approaches, see AI personalization.
Recommended products in ecommerce: where they win and lose
We have dozens of recommendation tests across carts, product pages, homepages and search. Here is what they show, by placement.
Blend CommerceBlend Commerce
DRIP Agency
Convertibles
Clean CommitIn the cart: the strongest placement, with one big exception
The cart is where shoppers have already decided to buy, so complementary add-ons fit naturally.
Blend Commerce added a recommendation module to a plant brand's cart page that previously had none.
Revenue rose 43%, conversion rate rose 7.46% and cart completion rose 39%.
Convertibles tested a cart drawer for an 8-figure dog treats brand with a "People also bought" row of two complementary treats, alongside a cleaner layout.
The winning version added an estimated $33,072 a month in revenue.
Taking recommendations out can also hurt. Clean Commit tested a minicart redesigned around the checkout button, with recommendations and chat removed.
Conversion rate fell 9.9%, revenue per visitor fell 18.2% and AOV fell 13.2% over 7 days.
Now the exception. Convertibles ran a strap cross-sell in a premium bag brand's cart drawer, picked from 30 days of co-purchase data.
Both versions lost: -12.4% and -10.3% revenue per visitor.
Basket size held, so the loss came from fewer people buying at all.
The likely lesson: when the add-on invites a second decision (which strap matches this bag?), it can stall the first one.
Simple, low-price, no-thought add-ons fare better. More cart tests are in cart upsell tests and on the cart drawers page.
On the product page: visibility beats volume
On product pages, the biggest wins came from making recommendations easier to see or easier to scan.
Blend Commerce moved PerTronix's recommendations above a long vehicle-fitment list, collapsing that list into an accordion.
Conversion rate rose 65.16% and revenue per visitor rose 131.35%, though significance was near 90%, below the usual 95% bar.
Conversion Fanatics cut a store's recommendations from 9-12 large images to five in a carousel without images.
Sales conversions rose 64%, revenue per visitor rose 151% and AOV went from $246 to $359. No sample size or significance was published.
Good Growth tackled the same problem for Pets at Home a different way.
Most shoppers never scrolled to the recommendations, so the team made adding to basket auto-scroll to them.
Clicks on recommended products rose 23% and AOV rose 1.2% in a significant 3-week test.
Ulta Beauty showed recommendations at the same moment, in an overlay after a guest added an item to the bag.
Revenue rose 9%, items per order rose 4.7% and conversion rate rose 2.3%.
Product-page recommendations do not always pay. Clean Commit added a same-category recommendation block to product pages: AOV rose 19.7% but conversion rate fell 9.9%, on 9,093 visitors.
Same-category items are alternatives, and alternatives can pull shoppers away from the product they came for.
Drukwerkdeal, a printing company, saw +5.6% completed orders after removing cross-sell links from product pages.
On sold-out pages: the most reliable win
A sold-out product page with only a "notify me" button is a dead end.
Similar-product recommendations give the shopper somewhere to go, and this is where related (not complementary) products earn their place.
- DRIP Agency added a "Similar Products" section for Oceansapart: +11.6% conversion rate and +21.7% revenue per user.
- Visually added similar products to The Oodie's out-of-stock pages: +9.5% conversion rate, with no sample size stated.
- Blend Commerce added an "Explore Similar Products" link on a plant brand's sold-out items: +4.59% conversion.
The same logic applies to search. An Optimizely customer sent zero-result searches to product recommendations and saw +2.74% revenue.
On the homepage and with recently viewed items
Homepage results are mixed. An Optimizely customer reported +136% clicks to product pages from personalized homepage recommendations.
But another Optimizely apparel customer's head-to-toe outfits carousel cut add-to-cart by 6.07%.
Clicks are not purchases, so judge homepage tests on revenue.
Recently viewed blocks help returning visitors find their way back.
Blend Commerce's recently viewed homepage carousel for Stone Creek Coffee showed +13% revenue per visitor.
Clean Commit's recently viewed section on product pages ran 35 days on 97,250 visitors and landed flat at +1.0% conversion with a 71.2% probability, a useful reminder that big early lifts often shrink.
How to test product recommendations
- Pick the job first. Complementary items for the cart and post-add moment; related items for sold-out pages and empty searches.
- Measure revenue per visitor, not clicks. Several tests above raised AOV while cutting conversion, and only revenue per visitor shows the net effect.
- Check mobile separately. In Clean Commit's same-category test, mobile revenue per visitor fell 4.7% while desktop rose.
- Watch for big-order noise. Very large AOV lifts on small samples often come from a few large orders.
- Keep the buy box clear. Recommendations should never push the add-to-cart or checkout button out of view.
For related tactics, see bundles, product page upsell tests, and segmentation if you want to show different blocks to new and returning visitors.
You can browse every result in the full test database.