How Schibsted’s AI model helped boost subscription sales

The system dynamically predicts what each reader is likely to engage with and serves tailored content to boost subscriptions sales and conversions.

By leveraging first-party demographic data and sales insights, this model generates real-time, on-demand recommendations without relying on traditional batch processing.

A/B tests showed a 75 percent increase in subscription sales from front-page articles compared to previous models, in some of the use cases, unlocking new revenue opportunities. Despite limited user data, this approach has proven highly effective, and Schibsted continues to refine it by adding more data sources.

"For any news site, driving sales is the lifeline of revenue growth, and our campaign takes an innovative leap to unlock its full potential," noted Christoph Schmitz, Product Manager, Schibsted Media & Tech, Curate. "The front page is the most powerful tool for this growth - it engages users already on our platform, requires minimal external cost, and delivers outstanding results."

Tackling the anonymous user challenge

Building on past insights, the team knew that delivering precisely tailored content to the right audience was key to engaging users and converting them into paying subscribers.

"The industry's biggest challenge in achieving effective personalisation lies in the lack of robust reading history and user data for non-logged-in users - an opportunity we've embraced to innovate and excel," Schmitz said.

During the past year, Schibsted has developed a machine learning model that seamlessly integrates diverse data sources to quickly identify which content types drive user subscriptions.

Leveraging existing data sources, such as age and gender predictions from its advertising business and processed sales data, the team built a compact set of features enabling precise, real-time content recommendations with minimal input data.

Real-time recommendations that adapt instantly

Unlike traditional batch processes, this system generates recommendations on-demand.

"This innovation ensures that recommendations improve instantly as new user data becomes available. No need for batch re-runs - our model adapts dynamically, factoring in contextual elements like time of day and day of the week to maximise sales impact," he said.

The model is fully integrated with Schibsted's broader content recommendation system (Curate), blending its insights with editorial signals and other performance metrics. This ensures optimal performance while maintaining strong editorial control.

"Given that we don't really have much data on our prospects for sales, being anonymous users, and we lack the lakes of data for all our users that our competitors in the app and social world can utilise, this is really about adopting the mindset of selling the strawberries you have at hand," Schmitz noted.

From 158 data points to 12 key features

During development, the team tested 158 data points for their impact on subscription sales, narrowing them down to around a dozen features.

While one could argue that segmenting content based on sales potential generalises both content and audiences, Schmitz believes the nuances in their approach challenge that perception.

The new model was tested against existing models, and positioned higher up on the front page to reach more readers. Against the older approach, front-page sales increased by an average of 75 percent.

Despite scarce and imperfect data for anonymous users, results have been "staggering," Schmitz said.

The team is now adding more data sources and exploring how the model can also optimise content for anonymous users to boost engagement, not just sales.

Model development and system integration

The implementation involved two main strands: Building the model and integrating it with existing systems. Schibsted's recommendation system already uses models for other purposes, so the new one fits well in its workflow.

Creating an online feature store capable of operating at scale proved to be a major software challenge.

"There simply weren't any off-the-shelf solutions on the market at that time," Schmitz noted. "For finding the right features to train the model on, we used all available user data we could get our hands on, and used a mathematical approach to determine what really mattered - what features are useful, and which ones are just noise."

Of the 158 data points tested, only a dozen were retained. The team is now adding new features and data points.

Developing at scale: Lessons learned

Models are relatively easy to develop, Schmitz said, but building the infrastructure to solve problems no one has solved before in a small team takes time.

When asked what he would have done differently, he said, "Set more developers on it."

Schibsted's in-house system combines static rankers and machine learning models, relying on features from its custom feature store to predict user preferences.

"We've harnessed the power of orchestration tools like Flyte and an in-memory database to ensure seamless operation of our Java-based internal application, powered by AWS DJL (Amazon Web Services Deep Java Library) for efficient model inferencing," Schmitz said.

The organisation is now scaling the solution across multiple newsrooms, and transitioning to Tecton - a managed feature store solution.

"By unifying feature availability and expediting onboarding for various brands, we're confident this change will further enhance user experiences across our newsrooms," he noted.

Neha Gupta

Multimedia Journalist

[email protected]