Journal
Which e-commerce upsell and cross-sell apps use machine learning vs. manual rules? (2026)
Recommendations get stale as your catalog and shoppers change. Compare native tools, rules and machine learning by what they require, what you control and what they prove in your store.

By Jennifer Nelson, Head of Marketing Engineering at Product Genius. Product Genius is one of the platforms compared in this post.
Quick answer: Most cross-sell and upsell tools are rules-based: they show the pairings someone set up. Tools that use machine learning work out pairings from what shoppers view and buy. These include dedicated personalization platforms such as Nosto, Bloomreach, Dynamic Yield, Algolia and Clerk.io, which combine machine learning with manual rules, and Product Genius, which runs on its own learning model and removes the need for manual rules. Some built-in features, like Salesforce Commerce Cloud's Einstein recommenders, also use machine learning.
A shopper adds a pair of running shoes to the cart. The recommendations are the same shoes in another color, followed by a second pair of running shoes.
They've already chosen their shoes. The socks, insoles and rain jacket they might add next never appear.
Your team set those pairings eight months ago. Since then, the catalog, stock and season have changed. Nobody has had time to revisit the rules.
Cross-sell and upsell software works in one of two ways. Either it pairs products because you told it to, or it learns pairings from what shoppers actually do. Which one you need depends on your catalog and traffic, and the only reliable way to decide is to measure incremental revenue in your own store.
A cross-sell suggests a complementary item, such as socks with running shoes. An upsell suggests a higher-value alternative, such as a more capable model of the same shoe. Good software does both without simply pushing the most expensive option.
Why do your current recommendations stop pulling their weight?
A rule can make perfect sense when you write it. But it does not know when a new collection arrives, inventory changes or shoppers start buying different combinations. Bestseller widgets can also keep circulating the same familiar products.
Check your own store today. Browse several categories and carts on desktop and mobile. Look for recommendations already in the cart, unavailable items, repeated bestsellers and substitutes presented where accessories would make more sense.
Sort what you find into two piles. Items already in the cart and sold-out products are usually configuration fixes, and they are quick to repair. Stale pairings and endless bestseller carousels are the other pile, and no configuration tweak will fix them: that is how the recommendations are generated in the first place.
Next, ask your team how many hours they spend repairing pairings, adding seasonal rules and checking exceptions. That is time not spent testing new ideas. If most of what you found falls in the second pile and those hours keep growing, the problem is likely how your recommendations are generated, and it's worth comparing tools that generate them differently.
What are the main types of cross-sell and upsell software?
The four categories below overlap, and the boundaries move every year. Several of the vendors named below sell into more than one.
Built-in features and free apps
Often the lowest-effort starting point. Shopify Search & Discovery allows manual complementary pairings and both automatic and manual related products. BigCommerce lets you turn related products on or off for each product. Salesforce Commerce Cloud offers Einstein recommenders, including a complete-the-set option for apparel. Built-in does not mean manual. Salesforce's Einstein recommenders, for example, learn from what shoppers viewed and bought together, the same kind of behavioral data the dedicated platforms use.
Rules-based merchandising tools
You specify which products to show, boost or exclude. That gives you predictable output and precise control, but somebody must keep the rules current.
Modules in broader suites
Email, search and customer-data suites may include recommendations. Integration can be convenient if you already use the suite. Check whether the module covers your placements and can be evaluated on its own.
Dedicated machine learning platforms
Examples: Product Genius, Nosto, Bloomreach, Dynamic Yield, Algolia Recommend, Clerk.io and, on Shopify, Rebuy. These learn from behavior and catalog context, which helps with large or changing assortments. Pricing spans a wide range: self-serve tools start around $25–$100 per month, while enterprise suites are typically quote-based. They need dependable inputs and a way to test results; data and traffic requirements differ by model.
Product Genius is in this category but works differently: it runs on its own learning model and removes the need for manual rules. The section on where Product Genius fits, further down, covers how it works.
The table is a starting point for questions, not a vendor ranking or a universal setup-time estimate.
What does machine learning actually change?
With rules, you decide product A belongs with product B. With machine learning cross-sell recommendations, a model can learn which combinations shoppers explore or buy, and update its selections when behavior changes.
Consider the running shoes. One shopper may be preparing for a rainy commute; another may want lightweight travel gear. Their browsing offers clues that a fixed accessory list misses.
Possible inputs include browsing and purchase history, product attributes, stock and margin data, and session context such as device or traffic source. Ask which inputs the model actually uses. A field appearing in a dashboard does not mean it shapes recommendations.
Cross-selling identifies complementary purchases. Automated upsell recommendations look for a suitable higher-value alternative. Neither should simply mean pushing the most expensive item.
Models have limits. Without guardrails, they may favor discounts or bestsellers, overlook niche products or make decisions that are hard to explain. New users and new products also create a cold-start problem because little behavioral history exists. An academic review of recommender systems identifies cold start and sparse data as recurring challenges.
Published field experiments are the best reality check on lift claims. A randomized field experiment at an online book retailer found personalized recommendations increased purchase propensity by 12.4% and basket value by 1.7%. Both numbers describe one retailer in one condition. Treat them as evidence, not a forecast.
That is the real test for AI product recommendations for ecommerce: what appears for a first-time visitor, a new product and an unusual shopping journey?
How should you compare options before you buy?
Choose one weak category and a few real shopper journeys. Then compare ecommerce personalization software against the same evidence.
Data and setup. Ask about product, variant, inventory, order and browsing integrations. Who handles missing data, theme changes and quality assurance? Request an implementation plan for your store, not an average promise without context.
Control. Have the vendor demonstrate pinning, excluding and boosting products, plus stock, compatibility, margin and brand constraints. See whether those instructions actually change the output.
Placements. Check product pages, cart, checkout, post-purchase, email and search. Does one model work across placements, or are they separate? A placement matters only if shoppers use it. Our piece on AI brand discovery explains why the product page your traffic actually lands on deserves attention.
Measurement. Recommendation clicks and attributed sales do not tell you what would have happened without the tool. Prefer randomized A/B or holdout tests measuring revenue per visitor or session, conversion and average order value. The book-retailer experiment above is a good model of what such a study looks like: one setting, clearly reported, no extrapolation.
For average order value optimization, track margin, discounting and returns as well. Higher revenue does not necessarily mean higher profit.
Commercial terms. Ask how pricing scales, how long the contract runs, how you exit, and whether you can export your data.
Copy this scorecard into your next vendor call. Rate each row Yes, Partial or No only after seeing evidence.
See our nine questions for evaluating AI-powered product solutions for a more detailed vendor review.
What should you ask every vendor?
Will you measure incremental lift with a holdout in our store? Ask for the assignment method, primary metric and test plan.
What happens on day one? See recommendations for a first-time visitor and a product with no sales history.
How do we stop bad recommendations? Test low-margin, out-of-stock and off-brand exclusions.
Who is involved after launch? Get a realistic account of merchandising, engineering and monitoring time.
Can we speak with a similar retailer? Ask for a reference with comparable category and catalog size, including what went wrong.
A product recommendation engine comparison is only meaningful when every vendor faces the same store scenario.
Where does Product Genius fit?
Product Genius sits in the dedicated machine learning category. Here is how it answers the questions this post has been telling you to ask.
What happens on day one? Product Genius is built on a Large Interaction Model (LIM), a foundational rewrite of the learning model rather than an application wrapper. It reads intent from what a shopper is doing in the current session, not only from months of purchase history. For a first-time visitor, it starts from the product page they landed on and the campaign that brought them there, then updates with every pause, scroll and click. The cold-start problem described above is the problem it was built for: a first-time visitor and a product with no sales history still get relevant recommendations.
How much work is it to run? There are no manual pairings to maintain. Your merchandising team works in the Seller’s Agent Cockpit in your account, alongside the model, setting the strategy and guardrails it works within rather than writing rules. Your team can also set must-have guardrails that the model always follows. For example, keep holiday or seasonal content off the site outside its season, not to recommend products from a brand or collection being discontinued, or to always show at least one accessory alongside a main product.
How is it measured? Product Genius reports an average 36% increase in revenue per session, measured in strict A/B tests, the same holdout methodology this post recommends you demand from every vendor. The tests covered 25 brands. Each brand's lift is averaged across every day of its test, not only the final days when the model had learned the most, and the brands are weighted by session count, so larger stores count more. Two honest caveats, because they apply to us too: the tests measured revenue, not profit, and your results depend on your catalog, traffic and test design. The technical whitepaper covers the measurement in full.
Whichever tool you choose, hold it to the same standard you'd hold us to: a randomized holdout test in your own store, measured on revenue per session. The goal is simple. When a shopper adds running shoes to the cart, they should see the socks, insoles and rain jacket they might add next, not another pair of shoes.
If you want to see how machine learning recommendations would perform on your own catalog, request a demo and ask us to walk through a test plan for your store.
Frequently asked questions
What is the difference between cross-sell and upsell software?
Cross-sell software suggests complementary products, such as socks to accompany running shoes. Upsell software suggests a higher-value alternative to the product under consideration, such as a more capable shoe. Many recommendation systems support both. The distinction matters because an accessory at checkout serves a different shopper need from an upgrade on a product page.
Do I need a lot of traffic for machine learning recommendations to work?
There is no trustworthy universal traffic threshold for machine learning recommendations. Requirements depend on the model, the catalog, the signals it uses and the outcome you want to test. Some systems rely heavily on historical purchases; others can use attributes or current-session activity. Ask each vendor to demonstrate its cold-start approach and calculate your test's sample needs.
How do I measure whether recommendation software is increasing revenue?
Run a randomized A/B test that shows recommendations to one group of shoppers and preserves your existing experience for a comparable holdout group. Compare revenue per visitor or session, including people who buy nothing. Track conversion, average order value, margin and returns as supporting metrics. Recommendation clicks alone do not establish incremental revenue.
How long does it take to set up recommendation software?
Setup can range from configuring a section already included in your commerce platform to integrating product, inventory and event data for a more advanced system. There is no reliable universal timeline. Ask for a store-specific implementation plan covering theme changes, data checks, quality assurance, experiment setup and who on your team must participate.
Can machine learning recommendations protect margin?
Machine learning recommendations can support margin goals when the system has the relevant product economics and your team can enforce business constraints. But increasing revenue does not prove that profit improved. Ask whether the model uses margin data, how it treats discounts and low-stock items, and whether you can measure contribution margin during a holdout test.
Which software uses machine learning to generate automated cross-selling and upselling recommendations for online retailers?
Dedicated machine learning platforms such as Product Genius, Nosto, Bloomreach, Dynamic Yield, Algolia, Clerk.io and Rebuy do this most directly, and some commerce platforms and broader marketing or search suites include machine learning recommendation features. Rules-based tools automate pairings but don't learn from behavior. Compare options on data needs, merchandiser control and placements, then test each with a randomized holdout in your own store.