What is the value of personalized recommendations?

The growth of ecommerce has created more competition online, and online sellers are always looking for something that will tip the scales in their favor. One way to gain that advantage is a personalized experience for each customer. To build it, you first need to know individual users and their habits. With a growing number of products, a growing number of users and a changing market, you cannot prepare a personalized experience by hand.

The biggest companies understood the value of personalized recommendations long ago and hired large teams of data specialists to develop and maintain their recommendation engines. Some companies, such as Amazon and Google, are so focused on their data that they can be seen as data science companies. Netflix gives the clearest numbers: according to a 2015 paper by two of its executives, recommendations influence about 80% of the hours streamed on the service, and a typical member loses interest after 60 to 90 seconds of searching for something to watch.

What is a recommendation system?

A recommendation system, or recommendation engine, is software that predicts which items a user is most likely to want and puts them in front of that user. It can personalize the content of pages for each user individually. Other channels, such as email newsletters or mobile notifications, can be personalized too. User interactions from many channels feed the recommendation system, which makes recommendations more accurate and the service more personal.

Data is crucial for good recommendations. You start with user profiles and the attributes of your products. You can begin with anonymous visitors to your site and remember their interactions. Once you recognize their identity, merge their profiles across all your channels. The most valuable data for recommendations is the history of users and their interactions with items. For new products, provide at least some attributes, such as a text description, to improve the recommendations.

Recommendations are one part of a wider personalization strategy, next to segmentation, dynamic content and triggered messages; for that bigger picture, see how personalization changes the face of marketing. Before a company decides to build and maintain its own recommendation engine, it should weigh the costs and benefits. Storing and maintaining such data is not cheap, but it can be used in many other ways.

Types of recommendation systems

There are two classic types of recommendation systems, and most real systems combine them:

• Collaborative filtering recommends what people with similar behavior liked. Its item-to-item form, described by Amazon engineers in a 2003 paper, looks at each item a customer bought and finds items that buyers of it were unusually likely to buy too. In 2017 the journal IEEE Internet Computing chose that paper as the one from its history that had best stood the test of time.
• Content-based filtering recommends items similar to the ones a user already liked, using their attributes: category, brand, price, color, ingredients or the words in a description. It works for new products that nobody has bought yet.
• Hybrid systems mix both and add context such as time of day, device, location, season and stock levels. Large stores and streaming services typically use hybrids.

The models behind these types have changed a lot. In 2009 Netflix paid a $1 million prize to the team BellKor’s Pragmatic Chaos for improving its rating predictions by 10.06%, yet it never put the full winning solution into production: it said the extra accuracy did not justify the engineering effort, and by then most members were streaming rather than rating DVDs. Today’s engines use deep learning on sequences of clicks in real time, and since 2024 Amazon has used large language models to write the recommendation rows themselves, replacing generic labels such as “More like this” with ones like “Gift boxes in time for Mother’s Day”.

Every system faces the cold start problem: with a new user or a new product there is no history to learn from. The usual answers are content-based rules, best sellers in the visitor’s category or region, and a short question at sign-up about what the person is looking for.

Personalized product recommendations examples

In ecommerce, recommendations appear at almost every step of the visit. The most common placements are:

• Homepage: “Recommended for you” and “Recently viewed”, so a returning visitor continues where they left off.
• Product page: similar items for people still comparing, and “Frequently bought together” for accessories and add-ons.
• Cart and checkout: complementary products that complete the order, such as batteries, a case or a matching item.
• Search results and category pages: the same results ranked differently for each shopper.
• Order confirmation and after the purchase: refills, care products and the next item in a series.
• Email and push: abandoned cart reminders with similar products, back-in-stock and price-drop alerts, a weekly digest of new items in the shopper’s favorite categories.

A few practical rules apply everywhere. Show a small number of items that fit the space without scrolling. Explain why an item is there (“because you viewed”, “popular in your area”), which makes the suggestion more credible. Do not recommend what the customer has just bought, unless it is a product they reorder. Refresh recommendations as stock, seasons and prices change.

Recommendations are also one of the easiest ways to make a campaign personal: the benefits of personalized campaigns come largely from showing each person the right product at the right moment.

Costs of a recommendation system

What costs do you have to cover? If you decide to develop a recommendation system in-house, you need data scientists who build prototypes, developers who write production code, and engineers who set up and maintain the infrastructure. When this post was first written in 2019, the development phase usually took 2-3 years and cost considerable money, depending on the cost of staff and hardware. On top of that come running costs (infrastructure and equipment) and the specialists you need to keep the algorithms good and improve the system, who are not cheap.

However, you can avoid most of these expenses by using an external recommendation system. In 2026 that is the easier route for most companies. Cloud services sell recommendations as an API with pay-as-you-go pricing: Amazon Personalize, generally available since June 2019, is based on the same technology Amazon uses in its own store, according to AWS. Customer data platforms and ecommerce platforms include recommendation modules, and smaller shops can add them as plugins.

Buying also helps with quality. The quality of in-house recommendations depends on whether the team has enough capacity and knowledge to develop state-of-the-art algorithms and run them at scale. Building your own makes sense mainly when recommendations are the core of your product, as they are for a streaming service.

Advantages of recommendation systems

Let’s start with the most general benefits. A recommendation system gives you a comprehensive view of your customer base and products. The user profiles kept in it are based on past interactions with items and allow advanced analysis in business reports and dashboards generated on a regular basis. Such reports can show problems early, so you can avoid them, and business decisions based on this analysis can save you money.

Recommendations usually shorten the time needed to find an item and make it much more likely that users discover other interesting things. The result is greater user loyalty and satisfaction. Users typically interact with more items, which leads to more purchases and higher profits. In addition, newsletters, personalized content promotion and push notifications encourage users to come back, increase how often regular users visit, reduce cancellations and raise customer value.

How to measure the value of recommendations

Many companies regularly define and evaluate their key performance indicators (KPIs), which makes it easier to measure the impact of a recommendation engine accurately. The measurement is done with an A/B test, in which users in group A get personalized recommendations, while group B gets standard recommendations such as best sellers.

Ecommerce is the domain where the benefits are easiest to evaluate. Revenue from personalized recommendations can be tracked with a token or simply by adding up the prices of purchased products that were recommended a few minutes earlier. However, that number does not reflect the fact that the customer might have bought the product another way, or bought something else without any recommendation. That is why the comparison against group B matters more than the attributed revenue.

The metrics worth tracking:

• Click-through rate on recommendation widgets and emails.
• Conversion rate of sessions that included a click on a recommendation, compared with the control group.
• Average order value and items per order.
• Revenue per visitor, the best single number for comparing groups A and B.
• Repeat purchase rate and, for subscriptions, cancellations.

Risks, and how to turn personalized recommendations off

Recommendations can also fail users. Systems trained only on clicks tend to push already popular items and show people more of the same, the effect often called a filter bubble. Suggestions based on sensitive purchases can feel intrusive. A good system mixes in some variety and new items and lets people say “not interested”.

Users increasingly expect control. Under Article 38 of the EU Digital Services Act, the largest online platforms and search engines must offer, for each of their recommender systems, at least one option that is not based on profiling. On YouTube, for example, users can pause or delete their watch history, after which future recommendations are no longer based on the deleted videos. Many stores, Amazon among them, let customers view and clear their browsing history. For a shop, giving people this control is a way to keep their trust rather than a loss.