Product recommendations

The value for business. Shedding the light on recommendation engines details.

Product recommendations in online stores are worth implementing for several reasons, as they can significantly benefit both the consumers and the retailers. Here are some key reasons why product recommendations are valuable:

  • Enhanced Customer Experience: Product recommendations provide a personalized shopping experience for customers.
  • Increased Sales: Personalized recommendations can lead to higher conversion rates and increased sales.
  • Time Savings for Customers: In a vast online marketplace, customers may feel overwhelmed by the sheer number of products available.
  • Improved Customer Retention: By offering a personalized and positive shopping experience, online stores can build stronger relationships with their customers.
  • Competitive Advantage: In the competitive world of e-commerce, providing a unique and personalized shopping experience can set a store apart from its competitors.
  • Data Utilization: Product recommendation algorithms leverage customer data to understand their preferences and behaviors.
  • Adaptability and Continuous Improvement: Recommendation systems can adapt and improve over time.

That is what GPT chat said and other sites quote. The fact that personalized recommendations are "worth implementing" seems obvious. It is much more difficult to answer the question "What is the real value of these recommendations - how much profit do they bring?"

What is the real value of product recommendations?

If we take the trouble to check out the results of such a query in Google, we will find out that 35% of Amazon's sales come from recommendations and Netflix achieved 75%. This information originates from some consulting company  - published a few years ago is like Garner's "80% of..." or has similar "trust me" certificate. It is undoubtedly impressive, which is why it is often quoted in many materials. Direct revenue increases are more often reported to lie between one and five percent, which can also be substantial in absolute numbers

It is undisputed that recommender systems can have positive business effects in a variety of ways. However, how large these effects actually are - compared to a situation without a recommender system or with a different algorithm - is not always clear.

Unfortunately, while nowadays a number of research datasets are available, they usually do not contain quantitative data from which the business value can be directly inferred. Furthermore, since the choice of a business measures is often specific for a domain, researchers typically abstract from these specific domain what gives simplified results.

The business value of the recommender systems is not adequately defined, measured, or analyzed, potentially leading to wrong conclusions about the true impact of the systems.

Companies usually do not publicly share the exact details about how they profit from the use of recommendation technology and how frequently recommendations are adopted by their customers – main purpose of such information is product or service marketing.

It is challenging to find universal recommendation system and success measures that correlate well with the different forms of business. We've all heard about The Netflix Prize - an open competition for the best collaborative filtering algorithm to predict user ratings for films, based on previous ratings. But the winning strategy was never put into practice - it was not clear if the potentially resulting increases in business value would justify the engineering effort to implement the winning strategy in a scalable manner. So their recommendation system keeps recommending you comedy if you've watched comedies before and encourages you to rewatch a movie you saw a few weeks ago.

Recommendations must have the right context

Overall, recommending weakly related products does not drive sales of the recommended products. Further, recommending popular products is not helpful unless the recommendations are relevant. So, popularity of recommended products is by itself not valuable. But if recommended products are both popular and relevant, that appears to deliver the greatest increase in sales of recommended products.

Recommending substitute products indeed cannibalizes sales of focal products in favor of the recommended products but net impact of recommending substitute products is positive - recommendations lower search costs and drive a significant increase in purchase incidences.

While tests might indicate that promoting already popular items is more beneficial than promoting long-tail items, the recommendation of (at least some) long-tail items might have direct or indirect sales effects in the long run. Such effects can, for example, occur when customers discover additional item categories on a shop through the recommendations over time or when customers later on switch to a paid version of a product that was originally recommended to them as a free trial. Increasing the diversity of the recommendation lists improves length of daily user sessions.

Analysis of product recommendations results expose that the gains do not accrue equally among products; some products with unfavorable locations in the recommendation network will lose from the widespread use of recommendation engines whereas products that are more strategically located in the network will gain substantially.

Product Recommendation Techniques

Recommendation systems often correspond to patterns of human behavior - people trust products and brands that have been bought by others, people like new products and opportunities. Hence the popularity of trivial types of recommendations - they always work. Product recommendations we encounter on websites are often filtered. Sponsored products, higher priced products, excess stock are mixed with native results. These are tips to enrich customer’s shopping experience, but above all, it's business.

Static recommendations

Not all product recommendation engines leverage complex algorithms to suggest what clients might have bought. Static recommendations are just what the name suggests - manually curated and selected specific products to be displayed as recommendations. These recommendations are often based on general trends, seasonal promotions, or marketing strategies. In this way e-commerce businesses have a certain level of control over the displayed products and can be particularly useful for highlighting new arrivals, clearing out excess inventory, or promoting specific campaigns.

Below some examples of commonly used static recommendation techniques found on major online stores.

Top Selling Products

Most customers purchase what other customers buy. It is highly likely that clients have very similar needs and when they see that others have purchased and enjoyed certain items, it instills confidence in the quality and popularity of those products. Displaying top-selling products provides social proof, indicating that many people have already made the same purchase. Why am I still not among these customers?

Latest Products

Great way to stimulate impulse purchases - new products often generate excitement and curiosity among customers. The human need to be cool, trendy and stylish lead customers to impulse purchases, especially if the products are perceived as novel or exclusive. Seasonal changes, holidays, or special occasions drives sales perfectly.

Category match products

When client buys a suit, it is very likely that he also needs a shirt and tie. And even if he already has them, it won't be inappropriate the store offers them (maybe customer can buy a better fit?) just in case. It also helps customers to explore products they were not aware of the full range of products available. This can lead to increased sales, cross-selling or upgrading product or services.

Where algorithms come in

In all of the above examples, "recommendation engine" is Excel or a simple database. Even if large in size. They cannot be underestimated - these simple methods of product recommendation are used by the largest companies of the globe. But when there is a need to get even closer to the customer's needs with recommendations, personalize the offer and hit the right moment then more sophisticated methods are required.

This provides an opportunity for presenting graph databases in action - they contain implementations of different types of algorithms to produce product recommendations.

Customer clusters [customer look-alike]

The objective is to identify customer clusters - groups characterized by similar purchasing patterns (although individual customers within each cluster may also make additional purchases). By effectively grouping customers into clusters, we can generate a list of commonly preferred items among them. Leveraging this information, we can tailor our recommendations to offer relevant products.

The initial algorithm we execute determines the similarity of customers by analyzing their purchasing patterns, as illustrated in image below [who bought what]:


The next algorithm, the Louvain algorithm, detects communities among clients taking into account their 'similarity' relations. So we have our clients belonging to two clusters; yellow and green.



Most commonly purchased items

What if we want to see what products are most common in the same shopping baskets? What does the 'similarity' of items look like - a customer who buys X probably also buys Y. This type of insight can be used for making recommendations to the customer. This task can be accomplished using one of the association algorithms, such as Apriori, or, preferably, its more recent and faster counterpart: FP-Growth

The graph database offers Node Similarity or simply query option. Below are five customer transactions presented as a graph.

And here's the result for top three pairs [client who bought X, also purchased Y]:

Bread => Eggs [2 instances]

Beer => Bread [2 instances]

Apples => Cookies [2 instances]

Similar products

There are at least a few good ways to find similar products. In a graph database, we can use the NodeSimilarity algorithm to analyze the properties of relationships or attached nodes [products have their own attributes and properties].

But let's turn our attention to vector search. Vectors are mathematical representations of data in a multidimensional space. In this space, each data has its own coordinates, and tens of thousands of dimensions can be used to represent complex data.

Imagine a cube with different types of fruit placed inside. In the figure below we have a color representation for six of them. But if we were looking for apricots, they would probably be somewhere next to nectarines as similar fruits. And that's how vector search works; everything that is in the dimension around the reference information is similar information (here, a product).


Many recommendation mechanisms are currently difficult or impossible to implement due to legal regulations, huge amounts of data, algorithm efficiency, implementation and operation costs - it takes a long time and deep expertise to build an effective recommendation engine. Some types of recommendations based on customer reviews and likes are susceptible to distortions. It is to be expected that recommendation engines will improve in the accuracy of the number of variables they take into account when calculating shopping suggestions - rapid development of computer computing power will overcome most of mentioned above barriers in the near future.


“Measuring the Business Value of Recommender Systems Research Commentary” (Research), “Measuring the Value of Recommendation Links on Product Demand” (Research), GPT Chat, Human Intelligence