A lot of people consider recommender engines as a tool to achieve cross-selling and up-selling strategies in ecommerce. However, these services can be more powerful and provide other functionalities that complement and enhance the final utility.
There are four key elements to consider in every good recommender engine: data, algorithms, design and metrics.
Data represents different properties and characteristics of users and products within the online store. They generally refer to actions that users perform on the products, which include visits to the product pages, shopping, ratings and comments, among others. But some of the main problems are data representation and collection. Representation can be qualitative (visits to product page, likes or dislikes, inclusion in wishlists…) or quantitative (mainly ratings). Collection can be implicit, like the data users provide for the simple fact of making the purchase process (e.g., visiting or purchasing products), or explicit, when users provide information in a proactive manner (e.g., ratings or comments). The latter is much more difficult to get, since the user must feel the need to provide it.
Algorithms use data to determine which products to recommend to each client on the web. Surely this is the most complex element, and the main entry barrier for a company that wants to develop its own recommender engine, not just by how easy or difficult it may be to develop these algorithms, but also for the scalability issues that occur as data size grows.
A positive point is that there are a number of well-known algorithms that produce good results for ecommerce. These algorithms generally use collaborative and content-based approaches, and they have proven results. The first problem here is that not all algorithms are suitable: the simplest produce worthless recommendations, and the most complex take too long to generate the results. Therefore, the algorithms used to get a successful outcome must sufficiently advanced, but also simple enough to make calculations in an acceptable time. The second drawback is that each algorithm has a number of attributes that make it more suitable for different types of business. So it would be very convenient to have a recommender engine that provides several algorithms and use the most appropriate in each section of our online store.
We know that design and usability in ecommerce are of great importance: a good product with a bad design goes unnoticed, and recommendations are no exception. Deciding the best location to show the recommender results, the number of items to display and the visible information for each product can encourage customers to follow a good recommendation. However, a bad combination of these may turn them invisible.
And finally we come to the metrics. Surely you will be familiar with the phrase “You Can’t Improve What you Don’t Measure”, and this is a truth that we should all take to heart in any website. It is also true for recommendations, either to check the improvements achieved by the system, or to compare the effectiveness of various recommenders and choose the best.
There are multiple recommender engines in the market that give support for cross-selling and up-selling in ecommerce, but most of them are too simple and offer only a minimal solution for a subset of the described key elements. If we want to be serious about our businesses, we should search for a tool that gives prominence to all of them.