E-commerce product recommendations: how they work and what good looks like

Imagine a shopper who has just bought a tent. They're getting ready for a camping trip, and there are still a few things they need.

The store's recommendation engine offers three more tents, the 3 tents they previously looked at.

Technically, nothing went wrong. The system recognized an interest in camping equipment, found similar products, and displayed them under "You may also like."

But the shopper doesn't need another tent. They need a sleeping bag, a lantern, and probably a few things they haven't thought of yet.

This is one of the fundamental problems with e-commerce product recommendations. Most systems are pretty good at finding products that are similar to what someone has already viewed or purchased. But similarity isn't necessarily the same thing as relevance.

The more interesting question is whether the recommendation helps someone make their next decision.

That's what separates a recommendation carousel that fills space on a page from one that actually helps people shop. And it's something e-commerce teams can measure more meaningfully than simply counting clicks.

What is the purpose of product recommendation for e-commerce?

At its simplest, a product recommendation is a suggestion about what a shopper might want to see or buy next.

But there's a difference between predicting what someone might click and understanding what they're trying to accomplish.

Think about a good salesperson in a physical store. Someone looking at tents might need help comparing sizes, deciding between materials, or finding the right equipment for a weekend trip.

The salesperson doesn't keep handing them tents indefinitely. They pay attention to where the customer is in the buying process.

Online recommendations should work the same way.

The job changes depending on where the shopper is:

• Homepage and new arrivals: Help people discover products or categories they might not have considered.
• Category and search pages: Help shoppers explore options, refine their preferences, and compare alternatives.
• Product pages: Help shoppers decide whether an item is right for them or discover something that might be a better fit.
• Cart and checkout: Help complete a purchase with relevant complementary items.
• Email and post-purchase: Give customers a useful reason to return based on what they've already purchased or explored.

Each of these jobs requires different logic.

For our camping shopper, recommending three similar tents might be useful while they're comparing options. Once they've purchased one, those same recommendations become much less helpful.

Now the goal is to complete the trip, not restart the shopping process.

This is why using the same recommendation widget everywhere on a site can produce disappointing results. The technology may be doing exactly what it was designed to do, but it isn't necessarily solving the right problem.

How recommendation engines work, in plain terms

There are several ways a product recommendation engine can decide what to show. Each has advantages, limitations, and situations where it makes sense.

Rules-based recommendations

This is the most straightforward approach.

A merchandiser decides that when someone views or buys Product A, the store should recommend Products B and C.

For example, someone buying a tent might see a sleeping bag and a lantern because the merchandising team explicitly linked those items.

The advantage is control. Merchandisers know what will appear, and they can make sure the recommendations reflect the brand's priorities.

For smaller catalogs, this may be all you need.

The challenge comes when the catalog grows or changes frequently. Maintaining thousands of relationships between products becomes a significant amount of work.

Popularity-based recommendations

These systems recommend products based on what's selling, trending, or receiving the most attention.

They're simple, relatively easy to implement, and can be effective when you don't know much about a shopper.

But there's a trade-off.

Popular products tend to get recommended more often, which makes them even more popular. Meanwhile, less visible products may never get a meaningful opportunity to be discovered.

And popularity doesn't tell you much about the individual shopper.

Collaborative filtering

Collaborative filtering uses patterns across shoppers to predict what someone else might like.

You've probably encountered versions of this approach under headings like "Customers who bought this also bought."

The idea is that people who behave similarly may share interests.

If thousands of customers buy a particular tent and then purchase the same sleeping bag, that relationship becomes useful information.

But collaborative filtering depends on having enough historical interactions to identify reliable patterns.

New products don't have a purchase history. New visitors may have no identifiable history at all. This is commonly called the cold-start problem.

Content-based recommendations

Instead of looking primarily at what other people bought, content-based systems look at the products themselves.

They might compare category, color, material, price, size, style, or other catalog attributes.

If a shopper likes a lightweight waterproof jacket, the system might recommend other products with similar characteristics.

This can be useful for new products without much purchase history.

But its effectiveness depends on product data quality. If the catalog has incomplete descriptions, inconsistent categories, or missing attributes, the recommendations have limited information to work with.

Hybrid and session-based AI recommendations

More advanced systems combine multiple signals.

They may consider purchase patterns, product attributes, current inventory, and what a shopper is doing during the current visit.

For example, a shopper lands on a product page for a winter coat. They look at several styles, linger on insulation details, skip expensive options, and return to a lighter jacket.

Those actions provide clues about what matters to them right now.

A session-based system can use that information to change what it recommends as the visit unfolds.

This is especially interesting because it doesn't necessarily require knowing who the shopper is.

The trade-off is that adaptive systems are more complex to evaluate and explain. Merchandisers need to understand what information is influencing recommendations and whether the results support their business objectives.

Ultimately, the algorithm matters, but so do the inputs.

Good catalog data, useful behavioral signals, and sensible merchandising constraints can matter more than having the most sophisticated recommendation model.

Where product recommendations go wrong

Most recommendation problems aren't especially mysterious. You can often find them by shopping your own site for a few minutes.

Some common examples:

Recommending something the shopper already purchased. A customer buys a product and immediately sees it recommended again. Unless it's a consumable or repeat purchase, that space could be used more effectively.

Confusing similarity with usefulness. Someone shopping for a dining table may need chairs, not six more dining tables. Similar products help with comparison. Complementary products help complete a purchase.

Ignoring availability. Recommending out-of-stock products creates frustration, particularly when the shopper has already demonstrated buying intent.

Optimizing for revenue without considering profitability. A recommendation that increases sales may be less valuable if it consistently promotes low-margin products or items with high return rates.

Assuming one click explains everything. A shopper clicks on a red dress. That doesn't necessarily mean they want every subsequent recommendation to be red. They may have been interested in the style, price, fabric, or something else entirely.

Showing too many recommendations. A page with multiple competing carousels can become difficult to navigate. More recommendations don't automatically create more discovery.

Some of these issues can be addressed with better algorithms. Others are simply merchandising problems.

A rule that excludes out-of-stock products may deliver more immediate value than an expensive new AI system.

The important thing is identifying which problem you're actually trying to solve.

What good looks like: how to evaluate your recommendations

One of the easiest mistakes is measuring recommendation performance only by click-through rate.

Clicks are useful, but they don't tell the whole story.

Suppose a new recommendation carousel generates 20% more clicks than the previous version.

That sounds promising. But what if those shoppers would have purchased anyway? What if they click more products but buy less? Or what if the recommendations simply redirect purchases away from other parts of the catalog?

A better question is whether recommendations are producing incremental value.

Measure what changes, not just what gets clicked

A randomized A/B test is one way to find out.

For example, you could divide eligible shoppers into two groups:

• A control group sees your current product recommendations, or no recommendations at all.
• A treatment group sees the new recommendation experience.

Both groups shop during the same period, under comparable conditions.

You then compare outcomes such as revenue per session, conversion, and average order value.

This helps distinguish between revenue associated with recommendation clicks and revenue the recommendations actually caused.

It's also important to choose a metric appropriate to the placement.

The metrics should reflect what that recommendation is supposed to accomplish.

Look at catalog coverage

Another useful measure is how much of your catalog ever gets recommended.

If a store has 5,000 products but the recommendation engine consistently surfaces the same 100, much of the assortment may effectively be invisible.

That doesn't mean every product deserves equal exposure. Some items will naturally perform better than others.

But poor coverage can indicate that your system is reinforcing popularity rather than helping shoppers discover products across the catalog.

Don't ignore business constraints

Revenue isn't the only consideration.

Inventory availability, contribution margin, return rates, and customer satisfaction all affect whether a recommendation is good for the business.

A system that increases average order value while promoting frequently returned products may not create as much value as the revenue numbers suggest.

Try the experience yourself

Not everything important appears in a dashboard.

Visit your store as three different shoppers: a first-time visitor arriving from an ad, a returning customer comparing products, and someone who already knows what they want.

Look at what gets recommended.

Does the experience adapt to what each shopper is doing? Do the products make sense together? Are the suggestions useful, or are they simply filling available space?

A few minutes of qualitative review can reveal problems that aggregate metrics hide.

A practical checklist for e-commerce growth teams

You don't need to rebuild your recommendation system to start improving it.

Here are six practical steps worth taking before investing in more sophisticated technology.

1. Give every placement a specific job. Decide whether you're optimizing for discovery, comparison, cart completion, or re-engagement. Choose one primary success metric for each.
2. Clean up your product data. Accurate categories, attributes, pricing, and availability make every recommendation method more useful.
3. Establish a baseline. Measure your existing experience before changing it. Test one meaningful change at a time so you understand what's driving the result.
4. Align merchandising rules with business goals. Decide which products should be excluded, which inventory needs attention, and where profitability matters more than volume.
5. Review different shopper segments. New visitors may behave differently from returning customers. Mobile shoppers may need different experiences from desktop shoppers. High-intent visitors may respond differently from casual browsers.
6. Keep reviewing performance. Product catalogs, customer preferences, inventory, and marketing campaigns change. Recommendations that worked three months ago may no longer be relevant.

It's also worth recognizing that A/B testing has limitations, particularly when experiments overlap, audiences differ, or results are interpreted without enough context. We've explored some of these challenges in our article on the limitations of A/B testing.

The goal isn't to run more experiments for their own sake. It's to learn which changes actually help shoppers and the business.

Where AI merchandising fits

Not every e-commerce store needs a dedicated AI recommendation engine.

If you have a relatively small catalog, stable inventory, and straightforward product relationships, rules-based recommendations or the tools built into your e-commerce platform may work perfectly well.

Adding another platform introduces complexity, and that complexity needs to earn its place.

The calculation changes as catalogs get larger, product assortments shift more frequently, and shoppers arrive from different channels with different intentions.

Consider a shopper arriving from an AI search result, another from Instagram, and a third from a paid search ad.

They might land on exactly the same product page, but they're not necessarily looking for the same thing.

Traditional recommendation systems often start with the product. More adaptive AI merchandising systems can start with the shopper's behavior.

That's the approach behind Product Genius.

Our AI Seller's Agent observes how shoppers interact with a page, including scrolling, clicks, and where they linger. It uses those signals to adapt the products and supporting content shown during the visit.

Instead of waiting for a shopper to search again or click through several categories, the experience can respond as their interests become clearer.

An example from Pier 1

One interesting example comes from Pier 1's experience with Product Genius.

Traditionally, a customer looking at candles might be recommended more candles.

But the Product Genius system identified another relationship: shoppers interested in a particular fragrance could also be interested in that same scent as a diffuser, room spray, or oil.

The relevant connection wasn't necessarily the product category. It was the shopper's interest in the scent.

Pier 1 reported a 28.8% increase in revenue per session on mobile and 25.7% on desktop, compared with its control group, during randomized A/B testing of the Product Genius experience.

These results reflect the broader AI Seller's Agent experience, not an isolated test of one recommendation carousel. They also don't establish what another retailer should expect.

But they illustrate something important: when recommendations respond to what shoppers actually care about, they can reveal valuable relationships that conventional product categories miss.

And that's a useful standard to apply to any recommendation technology, regardless of the vendor.

Better recommendations start with a better question

Product recommendations have become a standard part of e-commerce. But just having them on your site doesn't mean they're helping.

The objective isn't to show shoppers more products. It's to help them find what they need, compare the right options, and discover things they might otherwise miss.

Sometimes that requires better product data. Sometimes it requires smarter merchandising rules. And sometimes it requires an AI system that can learn from what each shopper is doing in the moment.

Whatever technology you use, the same question applies:

Did this recommendation help the shopper make a better next decision?

If you'd like to see how adaptive recommendations work with your own catalog, request a demo of Product Genius.

Frequently asked questions

What is the difference between personalized and non-personalized product recommendations?

Non-personalized recommendations show products based on general rules, popularity, or catalog relationships. Personalized product recommendations use information about an individual shopper's interests or behavior to choose what appears. Modern systems can also personalize within a single visit, using current interactions rather than relying entirely on past purchases or customer identity.

How much data do you need before recommendation algorithms work well?

It depends on the method. Rules-based and content-based recommendations can work with little or no shopper history, provided the product catalog is accurate. Collaborative filtering generally needs more interaction or purchase data. Session-based systems can use recent behavior, although they still need reliable signals to make useful predictions.

How do you test whether product recommendations are increasing revenue?

Run a randomized test comparing your recommendation experience against a control or simple baseline. Measure revenue per session, conversion, and other metrics appropriate to the placement. Keep the groups comparable, include enough traffic to make the results meaningful, and check profitability alongside revenue. Recommendation clicks alone don't establish incremental impact.