Denník N: Taking a second shot at AI-powered churn prediction

Project name: AI-powered churn prediction

Newsroom: Denník N (Slovakia)

Solution: Automated daily scoring model to rank subscribers by churn risk and trigger targeted retention tests

Organisation staff size: 101 - 200

Tech stack: REMP (Readers' Engagement and Monetisation Platform), machine learning prediction models

Impact: Established data infrastructure and automated segmentation; long-term A/B retention testing ongoing

Core lesson: Predicting churn is accessible, but designing effective tailored interventions takes time and careful testing


Denník N, Slovakia's journalist-owned daily, has built one of Europe's most successful subscription businesses. Through the JournalismAI Innovation Challenge, it returned to a problem it had once given up on: using AI to predict which subscribers are about to leave, and learning what persuades them to stay.

The problem: When the low-hanging fruit is gone

Founded 12 years ago in Bratislava by 40 journalists, Denník N was created in response to oligarchs owning Slovak media. Around 70 per cent of its income now comes from online subscriptions, and the company has been profitable for a decade, reaching about €7 million in revenue and almost €1 million in profit last year.

With 75,000 subscribers in a country of five million people, its conversion rate is among the highest in the world. “When you are that far, the low-hanging fruit is gone,” says Tomáš Bella, Denník N's Chief Digital Officer. “There are only two ways to improve revenue: either sell more or increase your retention.”

Retention also has a regional twist. Central and Eastern European readers are less used to recurring payments, so publishers there still offer one-off options. “You have to persuade a lot of people to buy not just once, but 12 times a year,” Bella explains. Lessons from the USA or Western Europe cannot simply be copied.

Denník N had been here before. About eight years ago, it tried to build churn prediction into its Readers’ Engagement and Monetisation Platform (REMP), the open-source reader revenue platform it developed and now shares with around 40 publishers worldwide. The module could not outperform simple algorithmic rules and became the only part of REMP the team abandoned. “We knew that we needed it, but we just didn't know how to do it,” Bella recalls.

Building the solution: Predict first, then intervene

The new project has two parts: identifying which readers are at high or low risk of churning, and testing interventions that might move people from one group to the other.

To build the model, the team looked back at historical data where the outcome was already known. “You try many different ways, and most of them don't work,” says Bella. “But in the end you discover that if you combine these 10 or 20 signals, you could have predicted with some probability that these people would churn.” The back-tested model now serves as an early warning system.

The analysis produced surprises: for annual subscribers, the variety of reading mattered, not just the volume. All else being equal, someone who reads five articles about Ukraine is at higher risk than someone who reads two about Ukraine and three about sport. Future recommendation engines will therefore not simply offer readers more of the same but will also aim to diversify the content they recommend.

In practice, the tech team builds the model, which sorts every user by churn probability each day into automatic segments, such as the 5 per cent most at risk. The marketing and product teams then decide what to do. Readers almost certain to resubscribe within seven days, for example, need no discount at all.

“Estimating churn is not so hard today with AI tools. The intervention is hard,” says Bella. A reader who stops reading may need a reminder, or may keep paying regardless and be best left alone. Choosing between an emotional appeal to support independent journalism and a discount also carries a cost: “If you offer a lower price to someone who should have just received an emotional email, that makes a big difference to the revenue.”

The team designed five experiments. The main challenge, Bella says, was time rather than technology. Unlike sales A/B tests, retention results can take months to appear. “If your hypothesis is not good, you will find out six months later that you didn't get any useful data,” he says. Having underestimated how long tests would take, the team refocused on segments likely to deliver results within the programme. Following a mentor's advice not to look at the data too soon, it will keep the experiments running after the programme ends.

The opportunities: Open-source tools for every publisher

Revenue results are not yet in, but the infrastructure is in place. The modelling and testing technology will be released as open source and integrated into REMP, allowing other publishers to score their own subscribers for free.

Identifying high-risk readers, Bella says, “is 100 per cent applicable” anywhere. How to act on it depends on each publisher's audience and local context, though Denník N plans to share what worked and what did not. If experiments fall short, the team will design new ones on the same infrastructure.

The JournalismAI Innovation Challenge made the difference. “It was like a big mountain in front of us, and we were a bit afraid to tackle it again,” Bella admits. Without this programme, he believes, development work on their project would have been delayed by another three years.

Lessons for newsrooms

  • Churn prediction is no longer just for the giants: Until recently, only the world's biggest publishers could build churn models. “Now it has become so accessible that many more publishers should start doing it,” says Bella.

  • Prediction is the easy part: The real work is matching the right message or offer to each segment; discounting for loyal readers simply loses revenue.

  • Plan for slow results: Retention tests take far longer than sales tests. Design hypotheses carefully, allow extra time and resist reading the data too early.

Explore case studies from the 2025 cohort

The 2025 Innovation Challenge Report, Prompting Sustainability, collects case studies from all 12 grantees and reveals their journeys, challenges, and opportunities.

Read about their experiences, discover best practices, and find inspiration in their innovative approaches to integrating AI into journalism.


The JournalismAI Innovation Challenge, supported by the Google News Initiative, is organised by the JournalismAI team at Polis – the journalism think-tank at the London School of Economics and Political Science, and it is powered by the Google News Initiative.