Product recommendation engine for ecommerce
How the ranking works, where to place it, what the numbers say, and how to compare vendors
Published · Updated
A product recommendation engine is software that decides which products to show each shopper, on every page of an online store and in email, based on what that shopper and other shoppers have browsed and bought.
It replaces a fixed, hand-picked list with a ranking that changes per visitor and per placement.
This guide covers how the ranking works, where recommendations belong on a store, what the published research says about the revenue they drive, and the criteria to compare vendors on before signing anything.
Three generations of recommendation ranking
Most vendors on the market today sell one of three things under the same name. Knowing which one you are being shown is the fastest way to tell two demos apart.
1. Rule-based
A merchandiser writes the logic: Same category, same price band, current bestsellers. Predictable, easy to explain, and every visitor sees the same row. It goes stale as the catalog changes, and someone has to keep maintaining the rules. Most platform-native recommendation features sit here.
2. Collaborative and content-based filtering
Collaborative filtering reads purchase patterns across all shoppers: People who bought this also bought that. It finds relationships no merchandiser would write down, and it struggles with new products that have no purchase history, tending to over-recommend bestsellers. Content-based filtering compares catalog attributes instead, so it works on day one for a new product but only sees what the attributes describe. Most engines blend the two.
3. Generative and model-based
The current generation encodes products and shopper behavior as vectors, and increasingly reads the product image and description text directly, so an item can be matched on what it actually is rather than on how it was tagged. Behavioral signals, catalog attributes and context all feed one ranking per shopper. The practical difference shows on thin catalog data: The engine can place a new product sensibly before anyone has bought it.
What the research says recommendations are worth
Two datasets carry most of the published evidence: A Salesforce study of 150 million shopping sessions, and Barilliance's reporting on its own platform data. The figures below are theirs, with the source named on each row.
| Figure | What it measures | Source |
|---|---|---|
| 7% of visits, 26% of revenue | Share of traffic and revenue from visits with a recommendation click | Salesforce, 150 million sessions |
| 4.6x | Conversion rate of shoppers who clicked a recommendation versus those who did not (4.3x on desktop) | Salesforce, 2017 |
| 10.3% higher | Average order value on visits including a recommendation click, rising to 15.2% on tablet | Salesforce |
| 24% more likely | Add-to-cart lift for shoppers who clicked a recommendation, 31% on tablet | Salesforce |
| 37% vs 19% | Return rate of first-visit shoppers who clicked a recommendation versus those who did not | Salesforce |
| Up to 31% of revenue | Ceiling for recommendation-attributed site revenue, with 12% of sales on average across customers | Barilliance, own platform data |
| 288% increase | Conversion lift after a single recommendation engagement, from a 1.02% baseline | Barilliance |
| 1.7x | Effectiveness of recommendation widgets above the fold versus below it | Barilliance |
| 10% to 15% | Typical revenue lift from personalization, up to 25% for the best executors | McKinsey |
| 35% of purchases | Share of Amazon purchases attributed to its recommendation system | McKinsey, 2017 citing 2013 data |
The Amazon figure is a decade old and reads as a directional estimate rather than a current benchmark. The full dataset, with context on each figure, is in our product recommendation statistics post.
Types of product recommendations
Similar products
Alternatives to the product being viewed: Same category, similar attributes, comparable price range. Helps comparison shoppers find the right option.
Frequently bought together
Products commonly purchased in the same order. A phone case with a screen protector, running shoes with performance socks. This is the primary cross-sell driver and the type with the most direct impact on average order value.
Trending and popular
Products with rising purchase velocity or view counts. Useful for new visitors where there is no behavioral data to personalize with, and effective in fast-moving categories like fashion and limited editions.
Personalized for you
Recommendations built from the individual shopper's browsing history, purchase patterns and real-time behavior. The strongest type for returning visitors, and it needs behavioral data from at least one session to begin working.
Recently viewed
Products the shopper has already looked at. Low algorithmic complexity, high utility, and it removes friction for shoppers comparing options across several sessions.
Contextual
Recommendations that shift with context: Time of year, day of week, or an external signal like the weather forecast. Requires the engine to take in data beyond purchase history.
Where to place recommendations (and why)
| Placement | Best recommendation types | Primary impact |
|---|---|---|
| Product detail page | Similar, frequently bought together | Lifts average order value by adding complementary items to consideration |
| Cart page | Complementary, add-ons | Increases items per order at the highest-intent moment |
| Homepage | Personalized, trending | Engagement, return visits, faster path to relevant categories |
| Category page | Popular in category, personalized | Product discovery, surfacing depth beyond the top sellers |
| Personalized, replenishment | Drives repeat purchase and reactivation outside the session | |
| Zero-result search | Popular, recently viewed | Recovers sessions that would otherwise exit empty-handed |
Visibility decides more than placement type: Barilliance measured above-the-fold widgets at 1.7 times the effectiveness of those below the fold. Click-through rates vary widely with design, product mix and traffic source, so set your own baseline in the first month and track lift against it rather than against industry percentages.
How to evaluate a product recommendation engine
Eight criteria separate vendors once the demo is over. Take them to every shortlisted engine and ask for the answer on your own catalog.
Integration with your platform.
Is there a maintained app or plugin for the platform you run, or does the engine arrive as an API your developers wire up? Ask what the installation touches, whether it affects page load, and who maintains the connector when the platform updates.
Relevance on your catalog.
Test with your actual products. Frequently bought together should show genuine complements, not other items from the same category. A vendor that will not run a trial on your feed is telling you something.
Cold-start handling.
What does the engine show for a product with no purchase history, and for a visitor with no behavioral data? Every engine has this problem; the question is what the fallback is and how quickly it hands over to learned ranking.
Merchandising controls.
Can you boost, bury, pin or exclude products in recommendation slots, by margin, stock level, category or brand? Can a merchandiser do it without a developer? This is where daily work happens, so judge the admin experience, not only the feature list.
Segment and per-shopper customization.
Does the ranking differ per visitor, or only per segment? Can the logic differ per placement, so a cart row behaves differently from a product page row?
Reporting on conversion and average order value.
The engine should report recommendation-attributed revenue and average order value against a no-engagement baseline, and separate direct from assisted revenue. The assisted figure is usually the larger one and the easiest to under-credit.
Cross-channel reach.
Can one engine power on-site and email recommendations from the same behavioral profile? Two systems with two profiles produce weaker results than one, and they double the integration work.
Cost relative to what you use.
Check whether pricing scales on traffic, catalog size, revenue or seats, what happens in your peak month, and what support and documentation are included rather than sold separately.
What changes by platform
Shopify
Shopify's built-in recommendations are rule-based and limited to related products within a collection. Third-party engines integrate through the Storefront API or an app. Look for one that uses real behavioral data rather than product metadata alone, and check theme compatibility and page-load impact before installing.
Adobe Commerce and Magento
Adobe Commerce ships Product Recommendations powered by Adobe Sensei, capable but tied to the Adobe ecosystem. Third-party engines usually offer more flexibility and often better relevance. Implementation runs through widgets or the API, so check compatibility with your theme and custom modules.
WooCommerce
WooCommerce includes minimal related-products logic by category and tag. The plugin ecosystem is large and quality varies widely. Prioritize plugins that use behavioral data over category matching, and measure the page-weight cost before committing.
Headless and custom storefronts
API-first engines give full control over placement, design and logic. You build the frontend and the engine supplies the ranking. This takes more development effort and returns the most flexibility and the best performance, with no third-party widget overhead.
How Hello Retail's recommendation engine works
Hello Retail is a European ecommerce personalization platform. Its Product Recommendations rank products per shopper from behavioral signals captured on the storefront, and each placement is configured separately, so the logic on a product page can differ from the cart or the homepage.
The ranking is powered by Product Intelligence, which enriches the catalog with attributes derived from product images and descriptions. That matters for the cold-start case above: A product can be placed on what it actually is before anyone has bought it.
The same engine runs on site and in email. Hello Retail integrates with Shopify through a published Shopify app, and with Klaviyo so that audience segments and behavioral signals flow into Klaviyo flows and campaigns.
Frequently asked questions
What is a product recommendation engine?
A product recommendation engine is software that decides which products to show each shopper, on every page of an online store and in email, based on what that shopper and other shoppers have browsed and bought. It replaces a fixed, hand-picked list with a ranking that changes per visitor and per placement.
What is the best product recommendation engine for ecommerce?
There is no single best engine, because the right choice depends on the store platform, the size of the catalog and whether recommendations need to run in email as well as on site. The criteria that separate vendors in practice are integration with your platform, how the engine handles new products and first-time visitors, merchandising controls, reporting on conversion and average order value, and whether the same behavioral profile powers site and email. Shortlist on those, then test on your own catalog before signing.
How much revenue do product recommendations drive?
In a Salesforce study of 150 million shopping sessions, visits where the shopper clicked a recommendation made up 7 percent of traffic and accounted for 26 percent of revenue. Barilliance, reporting on its own platform data, puts the ceiling at up to 31 percent of site revenue, with 12 percent of sales attributed to recommended products on average across its customers.
Do product recommendations increase conversion rate?
Yes. In the Salesforce 2017 study, shoppers who clicked a recommendation converted 4.6 times more often than shoppers who did not, and 4.3 times more often on desktop. Barilliance measured a 288 percent conversion increase after a single recommendation engagement, from a baseline of 1.02 percent for sessions with no engagement.
What is the difference between rule-based and AI product recommendations?
Rule-based recommendations follow logic a merchandiser writes: Show bestsellers, show products from the same category, show items in the same price band. AI recommendations learn from behavior, so the engine can find that shoppers who buy hiking boots tend to buy merino socks two weeks later, across categories and price bands. Rules are predictable and go stale as the catalog changes; a learning engine needs data before it is useful and improves the longer it runs.
Where should product recommendations be placed?
Product detail page and cart carry the highest intent, so they usually produce the largest lift per impression. Homepage suits personalized picks for returning visitors and trending items for new ones, category pages surface depth beyond the top sellers, and email carries personalization beyond the session. Barilliance found widgets placed above the fold were 1.7 times as effective as those below it, so visibility decides more than placement type alone.
How long does a recommendation engine take to go live?
It depends on the platform and how the engine is installed. An app or plugin on a hosted platform can be running in days; an API-based engine on a custom or headless storefront takes as long as the frontend work does. Ask the vendor for a time to first live placement on your platform and for what the engine shows before it has behavioral data to learn from.
See the engine run on your own catalog
Hello Retail's recommendation engine is powered by Product Intelligence and runs across web, email and search.
See Hello Retail recommendations