You have probably already been the target of several recommendation systems today without noticing. The list of suggested films, the products that appear under "customers who bought this also bought", the songs the platform chose for you, the articles the site proposed next — all are the fruit of algorithms trying to guess what you will want. Recommendation systems have become one of the most successful and omnipresent applications of artificial intelligence, and a huge part of what giant companies sell goes through them. The good news, which many ignore, is that this technology has stopped being exclusive to the giants: it is now within reach of companies of any size.
A recommendation system has a goal simple to state and hard to execute well: suggest the right product, content or action, to the right person, at the right moment. When it works, it is almost invisible — the customer feels the platform "gets them", quickly finds what they want and discovers things they did not even know they were looking for. When it fails, it is annoying or ridiculous, and undermines trust instead of building it. Understanding how these systems work, what makes them good and where they stumble, is increasingly important for any company that wants to use data to serve its customers better.
This article demystifies recommendation systems: the logic behind them, their limits, and how to think of them in the service of the customer and not just the sale.
The two great logics of recommendation
Behind most recommendation systems are two fundamental ideas, often combined. The first is recommendation based on the behavior of similar people: if many customers who behave like you liked a certain thing, you are likely to like it too. This approach — known as collaborative filtering — does not need to understand anything about the product itself; it learns only from behavior patterns. It is the logic behind the classic "customers who bought this also bought that".
The second idea is recommendation based on the characteristics of the items themselves: if you liked this product, here are others with similar characteristics. This approach looks at content — the genre of a film, the type of a product, the theme of an article — and suggests things similar to what you have already shown you appreciate. Each of these logics has different strengths and weaknesses, and the best systems combine them to take advantage of both.
The problem of starting from scratch
One of the biggest challenges of any recommendation system is informally called the cold-start problem: what to recommend to a completely new customer, about whom nothing is known yet? Without behavior history, collaborative filtering has nothing to work on. The same happens with a just-launched product that no one has bought yet. This problem of the beginning — of the customer with no history and the product with no interactions — is one of the reasons building a good recommendation system is more subtle than it seems.
The solutions involve not depending on a single logic. For a new customer, you can recommend what is popular in general, or ask them for some initial preferences, or use the little you do know about them. For a new product, characteristic-based recommendation lets you suggest it to those who liked similar products, even without its own history. Managing these difficult beginnings well is an essential part of a system that works in the real world, not just in theory.
The risks a good system has to manage
- The bubble: always recommending more of the same can corner the customer into a narrow bubble, showing them only variations of what they already know and hiding discovery from them.
- The popularity effect: already-popular items tend to be recommended more and become even more popular, while good unknown items never get their chance.
- Empty relevance: recommending the obvious — suggesting batteries to someone who bought a remote — looks like personalization but adds no value.
- Trust: a clearly bad or strange recommendation undermines the customer's trust in all the others, even the good ones.
Discovery matters as much as relevance
A common mistake when thinking about recommendation is to focus only on relevance — suggesting what the customer will almost certainly want. But if a system only recommends the obvious, it becomes useless: the customer already knew about that. The real value of a good recommendation often lies in discovery: showing the customer something they did not know but will love. It is this balance between relevance and surprise that distinguishes a mediocre system, which only confirms what the customer already knows, from an excellent one, which helps them discover.
Finding this balance is an art. Recommendations that are too obvious bore; too risky ones look like errors. A good system mixes the safety of the relevant with the boldness of discovery, giving the customer a selection that both understands and surprises them. It is this ability to expand the customer's horizons, rather than narrow them, that makes recommendation create genuine value instead of just pushing more of the same.
A concrete case
A mid-sized e-commerce company had a vast catalog, but customers almost always bought the same popular products that appeared on the home page. A huge part of the catalog — good products that customers simply did not discover — barely sold, not for lack of quality, but for lack of visibility. The company did not have the resources to compete with the giants in sophisticated technology, but realized that a recommendation system, even a simple one, could solve this problem. They implemented a recommendation that, from what each customer had viewed and bought, suggested relevant products from the vast catalog that would otherwise stay hidden. The care they took was to balance relevance with discovery: the suggestions were close enough to the customer's interests to make sense, but included products they would not have found on their own. The effect was twofold and virtuous. Customers started discovering products that genuinely interested them and that they would never have seen, which increased sales and satisfaction. And the company started selling a much larger slice of its catalog, giving an outlet to quality products that had previously been forgotten. It did not take the giants' technology — it was enough to use the data they already had to connect each customer to what they would probably want. The value came not from a more complex algorithm, but from solving a real problem with the right logic.
Recommendation in the service of the customer
Like all personalization, recommendation systems face a tension between serving the customer and serving only the sale. It is possible to design a system to push what has the highest margin, or to addict the customer to more of the same, and in the short term that may even increase numbers. But the best systems — those that create lasting value — are the ones that genuinely help the customer find what interests them, even if that leads them to discover rather than just buy more. A customer who feels the recommendations serve them trusts them and returns; one who feels manipulated stops believing in any suggestion.
This is the difference between a recommendation that builds a relationship and one that exploits it. And, since trust is the most valuable asset of any customer relationship, the recommendation that respects it ends up, commercially too, winning in the long term.
In practice
If your company has a catalog or an offering that customers explore, and you still communicate with everyone the same way, a recommendation system — even a simple one — can connect each customer to what they will probably want, increasing sales and satisfaction at the same time. You do not need the giants' technology; you need to use well the data you already have about your customers' behavior, with the care to balance relevance and discovery. Are you helping each customer find what interests them within your offering, or letting them get lost and always buy the same thing?
