For any news publisher, growing digital subscriptions is a top priority. The front page is a powerful tool for driving conversions; it’s where we engage millions of readers daily, making it the best place to surface content that leads to a subscription.

One of the biggest challenges in this space is personalisation for anonymous users — visitors who aren’t logged in and for whom we have little to no historical data.

Unlike social media platforms or app-based competitors that thrive on deep user profiles, news publishers often lack the data outside the ad space needed to make personalised recommendations for these readers.

Creating AI-powered content recommendations 

To tackle this challenge, Aftonbladet developed a machine learning (ML) model designed to predict which articles are most likely to result in a subscription.

Instead of replacing our current recommendation system, this model became a new signal to boost the right content within Curate — our in-house editorial front-page management and content recommendation system. Thus, there is no change in the way newsrooms work or select content; it simply provides better recommendations for the nonsubscribers we want to convert.

Here’s how it works:

• Seamless integration: The ML model functions as an additional signal inside Curate, improving our existing real-time, on-demand recommendation system.
• Leveraging first-party data: It uses the same demographic data we rely on for display ad targeting, allowing us to validate our own data as well as sales data, which is instantly available to the model as it is updated.
• Rapid prototyping and deployment: Thanks to in-house infrastructure, we can now prototype models for new challenges faster than ever before and quickly deploy them into production.
• Online serving for real-time recommendations: Our infrastructure enables us to make data available for online serving, allowing models to be used dynamically for real-time content recommendations rather than relying on static, precomputed predictions.

With this approach, we can now identify high-performing content more quickly and place it in automated front-page positions, where it has the highest likelihood of converting a reader into a paying subscriber.

Best of all, we only show these recommendations for our non-subscribers. Logged-in subscribers get recommendations based on their interests to increase their engagement, served by other ML models and approaches.

Key insights

The impact of this innovation has been remarkable. When A/B tested against our previous models, our new approach led to a 75% increase in sales from front-page articles.

Here are some additional insights from our development and testing:

Whilst we had a list of over 150 potential data points to explore, the reality is that we, like most other news sites, don’t have consistent or complete data for all users, especially not for anonymous ones. And we don’t force our users to log in like the new “media companies.”

This made many of those data points impractical to use in a model. Instead, we focused on the small set of features we could actually rely on, and it turned out to be enough. This project showed us that having a few high-quality signals can outperform having many incomplete ones.

This approach not only boosted subscriptions but also opened the door to optimising content recommendations for engagement, as well as other commercial models that benefit all users, not just those likely to subscribe.

Infrastructure is the true challenge. At the beginning of our project, there were no good alternatives to an off-the-shelf solution available for hosting scalable feature stores and live models.

By first investing in our own infrastructure, we proved the validity of our path before adopting shelfware as it became available and mature. We’ve developed an in-house system that significantly accelerates our ML development and iteration cycles, while also reducing costs. This also makes it cost-efficient, as we can switch out parts with shelfware as soon as they become available and stable.

What’s next

This success is just the beginning. We are already expanding and refining the model by integrating additional data sources and fine-tuning its predictive capabilities.

Moreover, we are now scaling this solution across multiple newsrooms, ensuring our approach to AI-powered recommendations delivers value across all our brands. By leveraging a unified recommendation system, we can streamline onboarding and enhance user experiences across our entire media network.

We have also expanded this setup for other commercial products that face the same challenge, converting engagement into commercial revenue, such as affiliate.

The news industry faces unique challenges in personalisation, particularly for anonymous users. But by thinking differently — leveraging first-party data, integrating ML recommendations into our existing systems, and prioritising real-time adaptability — we’ve shown even limited data can drive meaningful impact.

This innovation not only helps us sell more subscriptions but also sets the stage for more intelligent, more engaging content recommendations that benefit all readers. As we continue to refine our approach, we’re excited to see where this potential takes us next.

Banner photo by Emma-Sofia Olsson.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .