This article was originally published by The Fix and is republished here with permission. Learn about the latest from the world of European media by signing up for their newsletter.
Operating since 1860, Aftenposten is one of Norway’s largest and leading digital media houses, with over 250,000 subscribers.
Over the past years, the publisher realized that many of the news products available today don't align anymore with the digital habits of their consumers. For example, readers in the afternoon may see a different front page than those in the morning, which may alter their perception of the most significant articles of the day.
To counter this, they decided to reinvent their front page by using AI. Aftenposten's front page now consists of manually selected articles and articles chosen by an editor-controlled algorithm to provide the best combination of what readers should read and what they want to read.
Thanks to this, they can offer a balance between providing users with information they should be aware of regardless of their interests and encouraging them to return for more content that piques their interest by dynamically weighing editorial signals and personal interest.
The combination of pre-existing rankers and human-in-the-loop personalization resulted in a roughly 25% boost in subscriber click-through rates (CTR) for front-page items and a subscription uplift of up to 11%.
The Fix spoke with Hanna Lind Jørgensen, data analyst for Schibsted Media group, Aftenposten’s parent company.
How did the idea of homepage personalization come about?
Aftenposten has seen an ongoing shift of readership from print to digital, as more readers choose to access journalism online. A digital newspaper brings with it the opportunity to present stories in a more dynamic and adaptive way.
We were early adopters of automated front page features such as replacing articles that had already been read by a user or viewed a certain number of times. This allowed us to keep the page fresh and responsive to changing reader behavior.
We then conducted in-depth research in the newsroom on our user segments, and it showed that our readers tend to skew towards higher age groups. It also revealed that preferences around news consumption vary across age and gender. While some readers show a strong interest in politics and sport, others are more drawn to long-form opinion pieces or articles focused on health and well-being.
These insights showed the potential to further tailor the front page with the help of tools like recommender systems, shaping it to better reflect the diverse needs and interests of our individual readers.
How does it work? How do you adapt the content to different users?
Our personalization system is enabled by default for all consented users on the front page. However, to receive recommendations, a user must first reach a certain threshold of activity.
Users who qualify are shown recommendations based on their own reading behavior as well as the behavior of similar users, a method known as collaborative filtering. For example, if many users who read article A and article B also go on to read article C, the system may recommend article C to others who read A and B.
Can you share some concrete examples?
For instance, users frequently read articles about Taylor Swift (article A) and general pop culture (article B). Even if they don’t typically engage with sports content, they might still read articles about Travis Kelce (article C), not because of an interest in football, but due to the connection through Taylor Swift. So the model will try to suggest Travis Kelce articles to these users.
Another example, users read articles about summer travel delays (article A) and Donald Trump’s latest legal developments (article B). Even if they don’t usually click on articles about mountain hiking (article C), they might still recommend one simply because others with similar reading habits found it interesting.
The system learns from these patterns and can surface content that isn’t obviously related by topic but is relevant based on how users actually engage with the articles.
In addition to recommendations, we incorporate features such as removing articles the user has already seen or read and ranking each article based on factors like news value, time spent on the front page, and more. Together, this contributes to the final overall personalization.
Are the editors in the loop?
During the process of exploring new technology to gain a more dynamic front page, it has remained a priority to uphold the journalistic mission for Aftenposten. To that end, the internal personalization team, curate, and the product team have worked closely with the editorial department to ensure that the content served continues to inform the public as a whole.
This is achieved through a set of editorial rules; for example, journalists can assign a news value to each article, allowing those with higher importance to surface more prominently. In addition, certain positions on the front page, such as the top three slots, are kept locked/manually controlled to maintain editorial control and ensure that critical stories are always visible.
Has there been an increase in readers and subscriptions?
Over the past year, Aftenposten has seen a great increase in front-page performance for logged-in subscribers. Click-through rates (CTR) on personalized positions, which now make up over 90% of the front page, have grown by approximately 25%. This stands in contrast to 4% growth recorded the year before, when those same positions were not personalized. In parallel, we have also seen an increase in clicks per user (CPU) of 65%.
One notable effect of this personalization has been a shift in the types of content that surface. Through multiple experiment iterations, we’ve observed that more diverse topics are gaining visibility, particularly among readers aged 30–39.
And for the non-subscribers?
We’ve introduced dedicated sales rankers in specific positions on the front page for this segment. The rankings are designed to increase the likelihood of conversion by promoting content with a strong subscription-driving effect. Over the past year, engagement on these positions has increased by roughly 32%.
In controlled tests, we’ve seen subscription uplift of up to 11%, demonstrating the effectiveness of this targeted approach.
Have you since been contacted by other media outlets who would like to do the same thing?
Yes, we are in touch with various media outlets. More and more media organizations are working to make their journalism more dynamic and individually tailored by using tools such as recommender systems and also integrating artificial intelligence features.
Are you working on something even more advanced, such as more in-depth personalization?
We have been actively exploring new ways to make personalization even more dynamic and responsive to individual preferences. Our demographic studies revealed not only differences in topic interests but also in preferred content formats.
For example, some users show a strong preference for listening to articles, while others are more interested in video content. As a result, we are investigating how we might expand the availability of alternative formats such as audio and video.
In addition, we are exploring the potential for more user-driven personalization. Rather than relying solely on behavioral data, this approach would allow users to make active choices for instance, opting out of very Oslo-specific content if they don't live in the city in favor of other more relevant content.
Photo by Oliver Cole on Unsplash.
Read more articles by
