We've all lived this experience as customers: you go into an online store to buy one specific thing, and by the end of the process it's convinced you to also take home a second product you didn't even know you needed, simply because it showed up at the right moment under a "other customers also bought" banner. Amazon built a large part of its ecommerce dominance on this mechanism, but you don't need to be Amazon to apply it: today there are accessible tools for stores of any size that do exactly the same thing, with varying degrees of sophistication.
What a recommendation engine actually is
A recommendation engine is, in essence, a system that analyses behaviour (what a customer looks at, buys, combines) to suggest relevant products at the right moment, instead of showing the same generic catalogue to everyone equally. You don't need sophisticated artificial intelligence to get started: even the most basic forms of recommendation (for example, showing products from the same category, or best-sellers) already improve on showing nothing at all. More advanced systems use real behavioural data (what other customers buy together, what each specific person looks at) to fine-tune relevance far more precisely.
The most common types of recommendation
- "Frequently bought together" (cross-selling). Based on real purchase patterns: if 40% of people who buy a coffee machine also buy a pack of filters, showing that filter on the coffee machine's page makes mathematical sense, it's not a guess.
- "Other customers also viewed" (similar products). Especially useful when a customer is undecided between several similar options; it helps them compare without having to search manually.
- Recommendations based on personal history. If a customer has repeatedly bought products from a specific category, showing them new arrivals in that same category on their next visit or in an email is far more relevant than a generic email to the whole database.
- "Complete the look" or product kits. Especially powerful in fashion and home decor: showing how to combine the product being viewed with others to achieve a full set, not just a single piece.
- Recommendations based on repurchase timing. For consumable products (cosmetics, food, pet products), reminding the customer that they're probably running low on what they bought last time, at the statistically likely moment they're running out.
The real impact in numbers
The data varies by sector and execution, but figures commonly cited in industry studies put the impact of a well-implemented recommendation engine at a 10% to 30% increase in average order value, along with a noticeable improvement in the store's overall conversion rate, simply because the customer finds something they're interested in more easily. It's not a marginal effect: for many stores it's one of the growth levers with the best effort-to-result ratio, because once set up, it works automatically on every visit at no extra cost per customer.
How to start without a big-company budget
Both Shopify and WooCommerce have product recommendation apps and plugins accessible to small businesses, many with free plans or low monthly costs that already cover the basics (related products, frequently bought together, best-sellers). You don't need to start with an advanced artificial intelligence system: starting with simple, well-thought-out rules (showing products from the same category, or the most logical complements to each product configured manually if the catalogue is small) already produces a noticeable improvement over having nothing.
A specific case: the pet products store that raised its average order value
A pet accessories store with a catalogue of around 200 products added a "frequently bought together" block on every product page, initially configured by hand (manually pairing leashes with harnesses, feeding bowls with non-slip mats) before moving to an automatic system based on real purchase data. Average order value rose steadily over the following months, and the most interesting discovery was purchase combinations nobody on the team had anticipated (for example, that people who bought puppy toys very often also bought cleaning products for household accidents), information that later also helped improve the catalogue itself and the blog content.
The limit: when personalisation starts to feel creepy
There's a fine line between "this is useful" and "this is a bit unsettling," and businesses that cross it lose trust instead of gaining sales. Showing recommendations based on what the customer has looked at within your own store feels normal and useful. Using behavioural data from outside your store (for example, aggressive cross-site tracking) or explicitly referencing very personal data the customer doesn't understand the source of, generates distrust. The practical rule is simple: if you, as a customer, would find it uncomfortable for a store to know that about you, your customers probably would too.
Personalisation beyond the product page
A recommendation engine doesn't have to be limited to the product page. Applied to email marketing (showing different products based on each subscriber's history, not the same generic newsletter for everyone), to the homepage (showing relevant categories based on past visits), or even to the store's internal search results (prioritising what that specific customer is most likely searching for), the same relevance principle can extend to practically the entire customer journey, not just the moment of viewing a product.
The balance between variety and relevance
A frequent mistake when implementing recommendations is assuming that showing more options always helps. The usual evidence on purchase behaviour points the other way: presenting too many alternatives can cause decision paralysis, leaving the customer buying nothing because they can't choose between so many similar options. Showing three or four very well-chosen recommendations usually works better than an endless row of generically related products. The quality of the selection matters far more than the number of options shown, worth keeping in mind when configuring how many products appear in each recommendation block.
How to test whether a change to recommendations actually works
Before assuming a change to recommendation logic has worked (for example, switching from showing "similar products" to showing "frequently bought together"), it's worth testing it in a controlled way rather than rolling it out to the whole catalogue at once. A simple test with part of the traffic seeing the old version and another part seeing the new one, comparing average order value and click-through rate between both groups over a sufficient period, avoids the mistake of assuming a change improved things when the observed difference is actually due to another factor (a one-off promotion, seasonality) unrelated to the change introduced.
Recommendations and seasonality: adjusting for the time of year
A recommendation engine that doesn't account for seasonality can keep suggesting a winter-season product in the middle of summer, simply because the historical "bought together" pattern was calculated from data mixing the whole year. Adjusting recommendations for the time of year, prioritising current-season products over generic historical patterns, noticeably improves perceived relevance, especially in catalogues with clearly seasonal products like fashion, garden or sports items. Manually reviewing the recommendations shown when entering a new season, rather than blindly trusting the system to adjust everything on its own, remains good practice even with the most automated systems.
Common mistakes when implementing a recommendation engine
The first frequent mistake is activating the recommendations plugin and leaving it on default settings without ever checking what it's actually showing on each product page. Many tools, freshly installed and without enough sales history, display fairly absurd combinations (a product recommended alongside its own incompatible spare part, or two variants of the same item as if they were complementary), and those first few weeks of poorly tuned recommendations can create an impression of a careless store for the very first customers to see them, exactly when making a good impression matters most.
The second mistake is failing to distinguish between recommending products with similar margin and recommending with no commercial criteria at all. A system based purely on "bought together" can end up systematically suggesting the lower-margin product over an almost identical but more profitable alternative, simply because the cheaper one statistically sells more. Adjusting the system so that, among options of similar relevance, it slightly prioritises higher-margin products is a simple optimisation many stores never get around to configuring, unaware the option even exists.
The third mistake is not reviewing recommendations after removing products from the catalogue. When a product gets discontinued, any "bought together" block still referencing it on other pages breaks the experience and can lead straight to an error page, something that catches stores with regularly changing catalogues off guard more often than expected. A periodic check for broken links within recommendation blocks, though it may seem like a minor detail, prevents a quiet source of frustration that can be costing sales without anyone on the team noticing until they dig into the analytics closely.
The fourth mistake is not distinguishing between what works well on desktop and what works well on mobile. A recommendations block with six or seven products in a row can look perfectly fine on a large screen and become completely unmanageable on mobile, where the customer has to swipe a lot to see every option or, worse, doesn't even notice the block exists because it gets cut off right at the edge of the visible screen. Specifically checking how recommendations look on mobile, which usually accounts for most of any online store's traffic today, is a step many default setups don't cover well without manual adjustment.
A fifth mistake, already in the maintenance phase, is not reviewing recommendations when a completely new product arrives in a category never seen before in the catalogue, with no purchase history linking it to anything yet. During those first few weeks, leaving the recommendations block empty or showing barely relevant options is preferable to forcing nonsensical combinations just to fill the space, until the product itself accumulates enough sales history to generate reliable recommendations on its own.
Frequently asked questions
Do I need lots of customer data for a recommendation engine to work?
Volume helps, but it isn't essential from day one. You can start with simple manual rules based on product logic, and move to a system based on real purchase data once enough history has built up (usually a few months of sales already provide a reasonable base).
Is implementing product recommendations expensive?
The basic tiers are usually free or low-cost monthly within platforms like Shopify or WooCommerce. More sophisticated systems with advanced artificial intelligence cost more, but are rarely necessary until the catalogue and sales volume are considerable.
Can I use recommendations in email marketing too?
Yes, and it's usually one of the applications with the best return: instead of sending the same newsletter to the entire list, showing relevant products based on each person's history noticeably improves the click and conversion rate of those emails.
What if my catalogue is very small?
With small catalogues (fewer than 50 products), a complex automatic system doesn't add much over a few well-thought-out manual rules, because there's little variety of possible combinations. In that case, spending time deciding manually which combinations make sense is usually enough.
Can recommendations make a customer buy less, not more?
If poorly executed (irrelevant, repetitive, or intrusive), yes, they can create pushback or distract the customer from the product they already wanted to buy. The key is relevance: showing a few well-chosen options usually works better than flooding the page with generic recommendations.
How do I measure whether my recommendation engine is actually working?
Compare average order value and click-through rate on recommended products before and after implementing it, and also watch whether the percentage of orders with more than one product goes up. Most recommendation plugins include their own reports with these metrics already calculated.