To drive that effort, the data team worked on propensity to buy, next best action and churn models to reduce manual labor and incorporate data insights into their customer acquisition and retention journey.
This has helped them increase retention by over 14 percent in just three months, with a current subscriber base of 1.8 million across the Mediahuis group.
Mediahuis started its data journey six years ago and today has 40 people in the data team.
The publishing group has more than 30 news brands in Belgium, the Netherlands, Ireland, Luxembourg and Germany.
Jessica BulteData Science Business Partner, joined WAN-IFRA recently Digital Media Europe Conference in Vienna to talk about Mediahuis’ data-driven work over the last few years to achieve a more personalized customer journey experience.
Advantages and disadvantages of A/B testing
A/B testing plays a crucial role in understanding the best strategy to get anonymous users to sign up, convert and hopefully retain them, and that’s what Mediahuis has been doing for years, just like other publishers.
However, continuous A/B testing requires a lot of manual effort. The more complex the subscription economy becomes, with more opportunities for technology, the more manual work it requires, Bulte said.
“You’re pulling in different tools, learning from different places, and all of a sudden there’s chaos,” he said. “It’s also a slow process. Running 20 A/B tests at once isn’t efficient because you won’t be able to identify what worked.”
For example, a publisher might run an A/B test where test B performs better for 70 percent of its users compared to test A. So there is a clear winner.
However, a “winner-takes-all” perspective can be problematic, Bulte said. If a publisher uses the results of test B for its users, it ignores the 30 percent for whom test A performed better.
This directly impacted the brand’s existing challenge of trying to reach a younger audience. However, the majority of Mediahuis subscribers are older people.
“Going by Test B results will leave no room for personalization and differentiation. We would further target our optimizations to a larger segment of our audience,” he said.
Action prediction using propensity modeling
The next idea to explore was propensity modeling, which involves techniques for building predictive models that predict the likelihood of prospects buying a subscription based on their past behavior.
The bulletin interrupted the process of creating a predictive model.
- Data Collection (First Party and Behavioral Data)
- Clears that data
- Pattern recognition
- Making predictions
“Machine learning models are great for identifying past trends from data sets, learning from them, and predicting new patterns of behavior for new users,” he said. “This helps us segment that huge group of users into different subgroups: people who are most likely to buy, people who might buy, and people who have lost reason.”
A propensity model trained on data still requires manual effort, this time focused on differentiating between different audiences.
“This can again increase the chaos in all the different things you’re doing,” he said.
The key here, according to Bulthé, is not to have a predictive model, but one that can automate everyday tasks and improve the customer journey in a personalized way.
“If the trained model still requires you to do a lot of manual work and A/B testing and use that data to make decisions, then you’re data-informed,” Bulte said.
The solution that worked for Mediahuis was to move away from a data-driven way of working to a truly data-driven way, where testing is left to the algorithm.
Using past customer journeys
The team has now begun experimenting with modeling the next best course of action.
It involves the same initial data collection and cleaning steps. However, in this model, users are associated with campaigns that they have seen and/or interacted with.
This model learns from past customer journeys and predicts actions to take to optimize the customer lifetime value of micro-audiences based on several possible variables.
The next best action model is a step up from the propensity model because it informs the data team about the actions they need to take after a user subscribes, in terms of marketing.
“With this model, you can really personalize for different types of users. You know which campaigns to show them without manual effort because your data model has learned from previous customer journeys and is now ready to increase customer lifetime value,” Bulte said. “It’s also a great way to increase ARPU.”
The results speak for themselves
The churn model helped the team target groups of users and market their resources to those most likely to churn.
Users who showed a high propensity to deviate received marketing phone calls from Mediahuis. This increased retention by 14.17 percent in just three months.
Users with a medium likelihood of deviating were mailed a video of the editor-in-chief highlighting the importance of Mediahuis brands. This increased retention by 9 percent.
Work in progress
Over the next few months, the Mediahuis data team will further develop a model for next best practices.
The propensity to purchase model has been in production since March, and the next step is to conduct new dynamic payment A/B tests that will be ready for conversions over the next year.
