Denník N, a leading independent Slovak news site with around 70,000 subscribers, is utilising AI and machine learning tools to predict churn and retain subscribers

By: Veronika Munk, Denník N

Every subscription newsroom knows the feeling.

You celebrate a successful campaign, a major investigation, or a spike in new subscriptions. Then, a few months later, some of those readers quietly disappear. They do not complain, they do not announce their departure, they simply decide not to renew.

The problem is that we usually only know who has left. We rarely know who is about to leave.

At Denník N, one of Slovakia’s leading independent news organisations, this is more than a commercial challenge. Reader revenue is what protects our independence. Every subscriber who stays with us helps fund journalism that is free from political and business influence.

Like many publishers, we have spent years trying to improve retention. We have built renewal email flows, tested different messages, experimented with banners and offers, and learnt a great deal from A/B testing. But we have also reached the limits of what experience and intuition alone can achieve.

That is why our team joined the JournalismAI Innovation Challenge, supported by the Google News Initiative. Our project is built around a simple question: can artificial intelligence help us understand subscriber churn before it happens?

The surprising thing about churn

Publishers often think they do not have enough data. In reality, many of us have the opposite problem.

At Denník N, we have subscription history going back to 2017. We also have increasingly rich behavioural data: website visits, app usage, newsletter subscriptions, marketing interactions, and engagement with member benefits.

Yet many retention decisions are still based on relatively broad assumptions.

A reader who subscribed during a promotional campaign may have very different motivations from someone who has supported us for five years but has recently stopped visiting the website. One person might respond to a discount, another to a reminder of the value of independent journalism, and a third might simply need a better product experience.

Traditional retention systems often cannot tell these readers apart.

As a result, newsrooms tend to rely on one-size-fits-many solutions. We communicate with large groups of subscribers in similar ways because we lack a practical method for understanding individual risk.

We started wondering whether machine learning could help us see patterns that are hidden in the data but difficult for humans to detect.

AI for sustainability, not for content

Much of the conversation around AI in journalism focuses on content generation. How can AI help write articles? Summarise interviews? Translate text?

Our project comes from a different direction.

We are not using AI to produce journalism. We are using AI to support the business model that allows journalism to exist.

The goal is to identify signals that suggest a subscriber may be at risk of leaving. Instead of looking at one variable at a time, a machine learning model can examine combinations of behaviours: reading habits, subscription history, newsletter engagement, app usage, or previous marketing interactions.

The idea is not to let an algorithm decide what to do. The idea is to help humans make better decisions.

Internally, we focus on what we call sharp churn: subscribers who still do not have an active subscription thirty days after expiry. If we can predict the likelihood of sharp churn, we have a chance to intervene before the relationship is lost.

Learning before building

One of the most interesting lessons so far has been how quickly the conversation moved beyond algorithms.

Because we are building the project on Google BigQuery, experimenting with different machine learning models is relatively straightforward. We tested several approaches, but gradually converged on XGBoost, a model that is widely used and well proven for churn prediction and other forms of tabular data analysis.

What became more interesting than the choice of model was understanding what the model actually learns.

Like many newsrooms, we initially assumed that a handful of behavioural indicators would explain most churn decisions. The reality appears more nuanced. A small number of features account for a large share of the model's predictive power, but dozens of additional signals also contribute meaningful information.

Looking at feature importance gives us a much richer picture of subscriber behaviour. Early results suggest that the strongest signals come from several broad categories: payment and billing patterns, subscription structure and tenure, content engagement, and cancellation history. Together, these factors seem to capture not only how people use our products, but also how their relationship with the newsroom evolves over time.

This is one of the reasons we joined the Innovation Challenge. Building a churn model is ultimately not just a technical exercise. The challenge is translating machine learning outputs into insights that product, marketing, and audience teams can understand and act on.

In many ways, the project is as much about learning how to interpret subscriber behaviour as it is about training a predictive model.

Making AI practical

A successful AI project should not end with a dashboard that only data scientists understand.

One of our priorities is to make the output simple enough for product and marketing teams to use every day.

Rather than exposing technical complexity, we plan to translate predictions into clear risk scores that can fit into our existing CRM and audience systems. Teams will be able to identify groups of subscribers who may need different kinds of attention and design their own retention strategies around them.

AI does not replace human judgement.

It simply gives us a better starting point.

This is particularly important because we do not believe there is a single solution to churn. Some readers may respond to editorial recommendations, others to community benefits, others to a carefully targeted offer. Understanding the risk is only the first step; experimentation remains essential.

That is why controlled A/B testing will continue to be part of the project. We want to compare AI-informed interventions with our existing heuristic approach and learn what actually works.

Building something others can use

Our ambition goes beyond solving a problem for Denník N.

The project is closely connected to Reader’s Engagement and Monetisation Platform (REMP), the open-source audience and subscription platform developed by our team and now used by publishers in many countries. REMP was created to help independent media organisations build sustainable reader-revenue models, and we would like this project to follow the same philosophy.

If the churn prediction system proves robust and generalisable, we want to prepare it for future integration into the REMP ecosystem.

That forces us to think differently from the beginning. Which parts of the solution are unique to our newsroom, and which can become reusable tools for other publishers facing similar challenges?

We do not yet know the answer, but we believe that designing for openness will make the project stronger.

What success would look like

Success for us is not building the most complex machine learning system.

Success would mean that our product and marketing teams have better information, that our retention efforts become more personalised, and that we learn something meaningful about our readers.

It would also mean proving that AI can play a constructive role in journalism without touching the editorial process itself.

Our newsroom approaches AI with caution. We have clear internal principles around transparency, human responsibility, verification, and the protection of trust. Those values apply to audience projects as much as they apply to editorial work.

Ultimately, this project is about relationships.

Independent journalism depends on people who choose to support it. If we can better understand when those relationships are weakening – and respond in a more thoughtful way – we can help build a more sustainable future for our newsroom.

And if the lessons can eventually be shared through an open-source platform, perhaps they can help other independent publishers do the same.

_____

This article is part of a series providing updates from 12 grantees on JournalismAI’s Innovation Challenge, supported by the Google News Initiative’s second cohort. Click here to read other articles from all our grantees.

Can AI help independent newsrooms keep subscribers?