Technology has drastically changed the digital media landscape and how companies distribute news and engage with their readers.
The online shift has created a competitive, high-choice media environment where readers have more options than ever to reach information.
For media companies, it's no longer enough to create quality content and maximize visibility on the frontpage. In order to survive, you must find ways to increase user engagement and improve recirculation.
As such, many companies have embraced predictive analytics techniques and Artificial Intelligence (AI) to create a personalized experience for every visitor.
Companies like Buzzfeed and Apple News prove this method has been successful. It's clear personalization technologies are the new standard practice for increasing user engagement.
These companies leverage the predictive capabilities of AI and apply patterns about recent behaviors, views, and preferences for personalized recommendations.
But, AI algorithms pick up on patterns in real-time and make recommendations based on those specific patterns. This means there is often a natural tendency for recency bias to come into play.
Right now you may be thinking that doesn’t sound terrible. Timely recommendations are critical to creating a clear and unique path for each reader.
The problem is you may also be inadvertently excluding past content that could be relevant to your visitor's interests. Essentially, you are creating an echo chamber of content recency. And this can impact your user engagement and recirculation.
Why Relevance vs. Recency Matters
Producing content is at the core of what media companies do; therefore, they have years of quality content at their fingertips. However, because personalized content recommendations are based largely on current visitor behavior, a lot of this content may never be discovered.
Many media companies rely on generating traffic by leveraging trending content placed on the frontpage or section pages for maximum views. This often creates traffic spikes.
Predictive analytics tools then use the behaviors on these current visitors viewing recent articles and creates recommendations based on this data. This in turn creates an echo chamber. It only shows recent content, and it often doesn’t consider evergreen content that's been created in the past.
But, what if there is old content that is more relevant, just not recent?
This means readers probably won’t come across some of this old content that could be of interest to them.
For example, let’s say you have an article about the Royal family. Your visitor reads the article, but wants to get more information about something that happened in the past. If your recommendations only show recent articles, your visitors will never get to read that additional content. If they are searching for something specific, they will simply go elsewhere to find it.
Recency and topical content have their advantages, but they can also create limitations. Incorporating a wider set of evergreen content is a great way for companies to build more user engagement and increase recirculation. Access to evergreen content provides readers with information that is relevant to their specific interests, no matter when it was published.
So, how do you combat the built-in recency bias and incorporate more evergreen content?
Predictive analytics can reinforce the recency echo chamber problem, but with the right approach, they can also help you avoid it.
Using Machine Learning to Explore, Exploit & Avoid the Echo Chamber
At LiftIgniter, we use a different approach to Machine Learning to build our personalized recommendations and deliver the most relevant content.
We help media companies avoid the recency echo chamber by using a concept called explore and exploit. We recommend content that is related, but not necessarily the exact same subject, and see if there is a relationship there.
We continuously experiment with different content recommendations and if the content performs well, we will show it more often. If results are poor, we simply will stop showing it and will replace it with better performing content.
Instead of relying solely on metadata, we use Machine Learning techniques to determine intrinsic relationships. Metadata often makes recommendations by comparing apples to apples. The inferior metadata-based approach assumes that if two different articles are about dogs, they are related and should become recommendations. But this is not always the case.
With LiftIgniter, we put our beacon on a site and begin collecting information immediately. We then start showing recommendations based on the behavior of all of the users on the site. We analyze the paths they took, what they saw first, and how long they read an article. As long as one person has read an article, we track and analyze it no matter its recency.
Many personalization tools use a concept called collaborative filtering. Collaborative filtering takes the behavior history of what everyone else has done and says someone else will probably also want this.
Collaborative filtering often makes incorrect assumptions that can negatively impact the quality of your recommendations. For example, if a person reads articles A, B, and C, collaborative filtering assumes that an individual who only reads A and B, will also want to read article C.
However, this approach only considers the positive signals. At LiftIgniter, we have a lot more data that helps us determine intrinsic relationship statistics. Our method collects negative signals, meaning we collect what actions someone doesn’t take as much as what action someone does. There is a wealth of insight in this leftover data that we’ve highlighted in a past blog post.
We track this negative data by showing the recommendations several times. If the content is never clicked on, we take that knowledge and won’t make that recommendation again in that context.
By looking at both the positive and negative data signals, we have more control over the recommendations. We can explore content we've analyzed in the past that we think may have relevance even if it's not as popular.
Using this approach, we gain a deeper understanding of visitor behavior and we can provide more onsite discovery. This enables us to surface older content beyond what’s simply trending or is on the frontpage.
For media companies to stay competitive, increase user engagement and build loyalty, it will be critical to break the recency loop.