Of all news site readers, only a small number typically bother to register an account. And of all registered users, only a small number typically buy a subscription. So Scandinavian publishing house Schibsted is trying to use data to saving its marketing efforts — and subscription deals — for the readers who are more likely to pay up.

Schibsted’s subscription-purchase prediction model, developed by the company’s data science team, has been in use at four of the group’s Norwegian sites since last year: national newspaper Aftenposten and regional titles Bergens Tidende, Stavanger Aftenbladet and Fædrelandsvennen. The model, tested first at Aftenposten, predicts how likely readers who are already registered and logged into one of these sites are to buy a subscription, based on their browsing behavior and other activities.

“If we look back a bit, Schibsted has been in the print publishing business for more than 150 years, and we have operated free online news sites since 1995,” Eivind Fiskerud, Schibsted’s head of data and analytics for its Norwegian group, said. “It’s only the last four or five years that we have had user payments on our site. So this was a part of an effort to ramp up investment in digital growth.”

In addition to recording more reader behaviors, the prediction model gives the sites’ sales and marketing teams information they can use to target different groups of registered users with different digital subscription packages.

So far, the efforts seem to have paid off. Across the sites where it’s in use, the model has identified groups of readers 3× to 5× times more likely than average to buy a subscription. Sales staff at these news sites are then able to, for instance, target these specific registered users on Facebook with special subscription deals.

Data from Facebook campaigns showed that these targeted users were 22 percent more likely to subscribe when shown an ad linking to a paywalled article. Additionally, Schibsted’s marketing team spent an average of 35 percent less on Facebook advertising in trying to get each of these users to subscribe, compared to users that the model had pinpointed as less likely to subscribe.

Schibsted has always had a telemarketing strategy, calling registered news site users directly to offer subscription packages. In the dark about which users would be more likely to pay, the success rate for selling subscriptions over the phone was around 1 percent of all users contacted. When Schibsted’s marketers targeted the groups identified by its prediction model, that number rose to 6 percent.

Schibsted has the staff resources to devote to developing tools like its prediction purchase model. The company is one of Europe’s largest media groups, with leading news titles in Norway and Sweden, as well as extensive classified-ad interests in 22 countries and 7,000 employees around the world. The company has developed other tools focused on tailoring experiences to readers, such as one for automating the homepage story placement process, where user behavior such as clicks, conversion rates, and length of reading time factor into how stories are arranged on the site for the individual reader (a tool it has tested at one of its Swedish news outlets, Svenska Dagbladet, and at Aftenposten).

Norway’s Aftenposten has approximately 100,000 digital subscribers as of January. Digital subscriptions to the regional news titles Bergens Tidende, Stavanger Aftenblad, and Fædrelandsvennen bring Schibsted’s total count in Norway to 160,000. The company is aiming for 200,000 subscriptions across all these sites in 2018. (Norway’s population is just 5.2 million — making that scale all the more impressive.)

“What we didn’t really know before this project was how user behavior on the site relates to purchasing subscriptions. So that was unknown territory for us: What they are doing on the site, and what are the characteristics and patterns of those users who end up wanting to buy a subscription,” Fiskerud said. “Trying to crack that, and predict behavior, was the main business problem we had wanted to solve for.”

The prediction model is based on an algorithm that has been trained to identify the browsing behaviors of registered site users that go on to subscribe. It now takes between ten and 15 variables into account when determining a reader’s likelihood of subscribing, according to Ciarán Cody-Kenny, who worked on the model as part of Schibsted’s data science team.

Signals include straightforward factors such as a reader’s frequency of visits and the number of articles they clicked on, and other broader behaviors such as the number of devices used to access the site, past subscription history, and the proportion of weekend visits (which correlate with a higher likelihood to subscribe).

The data science team set the observational period, during which the model they were building analyzed registered users’ behaviors, at 14 days, after some experimentation. The data is stored as a shared list of user IDs and individual scores based on a user’s calculated propensity to take out a subscription, which the sales and marketing teams can then access and use in their own workflows.

“We started off with four weeks of data, but reduced to two weeks following a per-week analysis of some of the variables which showed that the fourth week — the last week — in the period had the strongest signal,” Cody-Kenny said. “Which makes sense, I think: Your most recent behavior is most predictive of your propensity to purchase next week.”

Ability to apply the prediction model across all Schibsted sites was an important aspect of the project right from the start, Cody-Kenny said. (So far, the company hasn’t implemented the prediction model at any of its other major news sites in Norway and Sweden, including Verdens Gang and Aftonbladet, one of the biggest news brands in Sweden, both of which combine free and premium access journalism.)

“Being able to replicate the process across these publishers was quite important for us. We approached this with a ‘nail it, then scale it’ attitude,” he said. “It’s somewhat inefficient if each of our publishers builds their own process from scratch. If we can build it once and prove the value for one site initially, then easily roll it out to others, we are saving on effort.”

The success of the prediction model so far has drawn some attention in the industry, and other media groups have asked about licensing the tool. For now though, selling tech built in-house commercially, à la The Washington Post with its Arc Publishing system, is not on the immediate horizon for Schibsted, according to Fiskerud.

“We’re focused on learning and using it to drive our own subscriptions and growth,” he told me.

The company still needs to collect more data on its overall effectiveness. But Fiskerud said that in-house technologies, built by and for news organizations, can help newsrooms gain more useful insight into audience behavior at a time when any owned data can be critical for publishers — even if this may be a level of marketing-data sophistication that other sectors, such as e-commerce, arrived at earlier.

“The importance of having control over our own data, and knowing who the user is, and building a deeper relationship with our subscribers is really key to surviving in journalism,” he said.