News companies constantly want to draw reader attention to relevant articles — in newsletters or on the Web site, for example. Push notifications are one of the most powerful ways to do that.

Sending a notification to hundreds of thousands of people is often the way to go to boost an article. At the same time, it is also risky: Because these notifications are intrusive, people quickly drop out when they get one that is not relevant to them. For breaking news, the risk of “annoying” users is small. However, for specific articles, it is important to get your audience right. With the help of personalisation via AI, it is possible to counteract that churn rate for notifications.

Look at people’s reading behaviour

One of the publishers using this strategy is Mediahuis, the leading publisher in the Low Countries. The company owns newspapers such as Het Nieuwsblad, which more than a million people read daily, and several popular regional titles. Instead of pushing all “non-breaking news” to millions of users, it selects key interested users.

“There are various ways of personalising push notifications,” said Jonas Boonen, head of product at Froomle. “All too often, these start from explicit preferences. In that case, people have to select which notifications they do and do not want to receive. It’s the first step, but ideally, you combine explicit preferences with implicit behaviour.”

That’s why push notifications use the same parameters as those used for personalisation in newsletters or Web sites. These are determined by looking at people’s reading behaviour.

In the case of Mediahuis, we look at which articles someone read over the past 90 days. When it comes to a regional article, we combine explicit and implicit preferences. People can subscribe to a specific region, but it’s also important to consider whether someone reads a certain number of articles about a particular area.

Moreover, user annoyance is minimised by limiting the number of messages each user can receive based on an agreed cap, like five push notifications per day or one per hour. In this way, we create segments of users who are potentially interested in a specific article. The algorithm only needs a few days to be fully up and running, and afterwards it gets smarter by taking into account more data.

Some companies, like Mediahuis, use categories and tags, but content can also be analysed using natural language processing (NLP). This allows for the enrichment of metadata.

Because the newsroom continues to use the same applications, their process doesn’t have to change.

Those who work with AI-powered personalisation companies and want to send a push notification should be able to choose how many people should receive it. They can also set other parameters by specifying how many relevant articles those people should have read. For regional articles, for example, publishers can choose whether they should also target neighboring municipalities.

Results of personalised push

The result? People get push notifications they care about, and churn rate (i.e., users opting out) decreases. Our research shows:

• The average churn rate for notifications (that is, daily users who unsubscribe compared to the total subscribed users) is normally around 1.5%.
• In the Mediahuis case, it’s 0.25% for users who receive personalised pushes and practically non-existent (as low as 0.02%) for users who click on personalised pushes.
• On average, we’ve achieved a click-through rate that’s 10 times higher with the personalised push notification compared to general news pushes for Het Nieuwsblad.

This is also the ideal way to hype an article. Like the TikTok algorithm, publishers can serve articles to a small audience first, and when it’s popular there, scale it to a broader audience.

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