Uberflip's Recommendation Engine Explained


Want to learn more about how Uberflip's recommendation engine figures out what to recommend to your Hub and website visitors? Here's how it works.


What is the recommendation engine?

The recommendation engine powers our recommendation systems, Content Recommendations and Site Engager. These systems create personalized content recommendations for visitors to your Hub and website, and the engine figures out what they should recommend to each visitor.

The old-fashioned way of recommending content is to simply offer up a list of popular or recent articles. The recommendation engine creates personalized recommendations that take the particular interests of each visitor into account. Compared to generic recommendations, this greatly increases the chances that a visitor will continue to consume your content and engage more deeply with your brand.


If you're interested in using Content Recommendations or Site Engager, you're probably wondering just how the recommendation engine works: how does it know what to recommend? In this article, we'll explain exactly how it comes up with recommendations.

Recommendation availability

Recommendation tools are not available on some pack types. For more information about your pack type and what's included, please contact your Customer Success Manager, or Uberflip Support.


A quick primer on recommendation engines

Here's a quick primer on what recommendation engines do, and how they do it.

What recommendation engines do

A recommendation engine is software that makes recommendations. Sounds simple, but "making a recommendation" is actually a pretty complex problem for a computer. That's because a recommendation is fundamentally a prediction: if a recommendation is the item a person would like the best among many options, pinpointing it means you need to predict how that person would rank all the available options.

A computers doesn't innately know how to make a prediction. To a computer, making a prediction is like any other problem: it can't solve the problem unless it's given a set of instructions designed specifically to solve that problem. Ultimately, this is what recommendation engines do: they give computers the tools they need to make predictions, and therefore recommendations.

How recommendation engines work

The technical term for "a set of instructions to solve a problem" might sound familiar: algorithm. In the context of recommendation engines, the algorithms that generate recommendations are commonly referred to as recommenders.

There are many different ways to make a recommendation, and each recommender represents a particular approach, such as "visitors with similar interests also viewed these items", or "these items are similar to the one you're viewing". No single approach will work for everyone, so recommendation engines typically rely on collections of many different recommender algorithms to generate their recommendations. This way, they can cover all their bases: instead of trying to predict a single best recommendation, they can simply generate a list of recommendations in various ways and let the user decide which one they like best.

Viewed as a complete system, a typical recommendation engine works like this: to begin, information about the user's preferences and the available options is fed into the recommendation engine's various recommender algorithms. Each recommender then processes this data in its own way, and serves up the item it has determined to be the best fit for the user. Once all the recommenders have run, the recommendation engine collects their output and applies any necessary refinements, such as removing duplicates. The end result of this process is a list of items that have been chosen to appeal to a user's interests: in other words, recommendations.



Recommendation engine: The basics

At the heart of Uberflip's recommendation engine are over twenty separate recommenders (algorithms) that each take a different approach to recommending content based on visitor interests, content similarity, and other factors.

Each recommender identifies several options, ranks them according to relevance, then presents just the top option as its "pick". This top option must also exceed a minimum confidence threshold to be presented — in some instances, a recommender can't be sufficiently confident that any of the options it has generated will be a good fit for a particular user, in which case it won't present one.

Because our recommendation engine filters recommendation options at the algorithm level, it can ensure that all of the options that it can select from are essentially equally good. As a result, the process of choosing the final recommendations is straightforward: the engine simply removes any duplicates, then randomly chooses the specified number of recommendations and presents them to the visitor. (On top of this, it also keeps track of visitors to avoid recommending something they've already viewed, and makes sure that a visitor doesn't see the same content recommendation more than five times.)


Recommenders: The engine algorithms

In this article, we won't dive into individual recommenders one-by-one, because that level of detail isn't very interesting. Instead, we'll look at the general concepts that the recommenders are based around. All of the recommenders rely on combinations and variations of these central ideas to generate recommendations, so this should give you a good idea of how they work at a high level.

Intent data

The best recommendations are specifically tailored to an individual's preferences. To do that at scale, Uberflip's recommendation engine relies on intent data. Intent data is information about the topics visitors are interested in — derived from actual observation of their browsing behavior — and focused specifically on the intent to make a purchase related to those topics. This makes it ideal for marketing-specific personalization, which is why about half of engines recommenders take intent data into account in some form.


The recommendation engine sources intent data from our partner Bombora, a leading provider of B2B intent data. Bombora uses machine learning to identify page topics based on keyword association, and tracks visitors across a large network of websites. Using this data, Bombora is able to build a profile of a given visitor's interests based on the topics of pages they visit. When you integrate your Hub with Bombora (which is a requirement for using the recommendation engine), the engine uses this intent data to understand both the topics that your Hub visitors are interested in, and the topics that your content is about.


Intent data is what gives the engine the ability to make truly personalized content recommendations for almost any visitor. For example, picture a user who is very interested in ABM. A recommender based on intent data will know this about that user, and can use this knowledge to look at all the available content that it has identified as relating to ABM. From there, it can rank all the options according to their relevance score, then recommend the item with the highest relevance score — an article that it has determined is ~92% about ABM, say.

Collaborative filtering

Collaborative filtering is a method used by most recommendation engines in some form, based on the idea that users with similar interests are likely to prefer similar items. Therefore, if you analyze the preferences of a large number of users, you can predict how any given individual would rate a particular item based on how similar users rated it. For example, say Amanda, Ben, and Charlie are all interested in content marketing: if we know that Amanda and Ben are also interested in ABM, then we can assume that Charlie is likely interested in ABM, too — even if he hasn't expressed this interest yet.

Collaborative filtering is one of the most effective ways of personalizing recommendations, so a large number of Uberflip's recommenders are based around it. And since collaborative filtering requires data about user preferences, most of these recommenders also rely on intent data to calculate user similarity and predict content preferences.

The recommendation engine's collaborative filtering (CF) recommenders can be broadly divided into two categories, user-based and topic-based:

  • A user-based CF recommender looks at a user's intent data to identify the topic they're most interested in, then recommends items that other users with the same top interest viewed most often.
  • A topic-based CF recommender looks at the current page a user is viewing to identify its topic, then recommends items that other users with this topic as their top interest viewed most often.

One particular strength of collaborative filtering recommenders is that they are self-refining, improving as the body of data that feeds into them grows. Plus, they are capable of generating recommendations that are not necessarily related to the topic the user is currently viewing, which can encourage content discovery and drive greater engagement.

Proximity and popularity

Lastly, a number of Uberflip's recommenders revolve around the concepts of proximity and popularity.

Proximity recommenders are based on the simple but effective idea that a user who is viewing a given item is probably interested in that topic, and would likely be interested in similar items about the same topic. By analyzing the content of all the items that can be recommended, proximity recommenders are able to calculate a numerical value for the "distance" between any two items based on how similar they are (i.e. how much overlap there is between the topics that they cover), then recommend the item that is closest (most similar) to the one currently being viewed.

Popularity recommenders have an even more straightforward premise: the most popular items are top performers because they have wide appeal among the target audience, and are therefore a safe bet to recommend to any user. This kind of recommender will identify the content with the most page views in a given period, or the highest click-through rates, and recommend it regardless of the user's current context. Although this type of recommender offers the least amount of personalization, its recommendations still tend to be effective. Moreover, it also serves a valuable role as a backup in cases where not enough intent data exists to offer more personalized recommendations.


Next steps

Now that you know how the Uberflip recommendation engine comes up with recommendations, try it out for yourself: learn how to set up Content Recommendations and Site Engager to begin serving algorithm-powered, personalized content recommendations at scale.

2 out of 2 found this helpful