While digital attention gravitates toward increasingly personalised experiences, most news organisations use the one-size-fits-all broadcast model to safeguard their journalistic mission. After seven years of experimentation with editor-in-the-loop news algorithms, Schibsted found we can have our cake and eat it, too.

Many of today’s news products don’t match their users’ digital habits: An afternoon reader sees a different front page than a morning reader, changing what they perceive as the day’s most important stories. Since most users only open three to five articles per day, this model risks exacerbating the gap in what our users read and, potentially, how they see the world. 

In the days of print, the newspaper remained static for 24 hours, and readers saw the same stories regardless of when they read it. Today’s one-size-fits-all front pages don’t fit users’ vastly individual digital habits. Our solution is a product that adapts to these differing habits.

The editor-in-the-loop

In Schibsted’s subscription newspapers, we use a combination of manually selected articles and algorithm-driven ranking of stories on the front page. The top three to six slots on the front page are manually chosen based on our editorial principles. An algorithm selects the other articles to ensure readers catch important news and see what other readers engage with.

Over the past year, we incorporated more personalisation into our ranking system. By having a dynamic weighting of editorial signals and personal interest, we can balance showing users what they should know regardless of interest while keeping them coming back to find content they have a particular interest in.

A dynamic front page with algorithmic ranking (which understands the vastly different behaviours of each user) lets us responsibly match niche articles to readers who are most likely to be interested in them.

Unlike algorithms on social media platforms, ours help rank a list of articles below the manually chosen top stories, all from a responsible publisher who has carefully reviewed the content and set the correct relative values for editorial metadata.

Our algorithm rules aim to create long-term value for our users, not to maximise short-term clicks or time spent. This means personalisation doesn’t just mean “more of what I like;” it means the content is tailored to the user, while what it is optimised for depends on our goals.

Choosing news over personal interest

The editorial metadata we incorporated is crucial for making the algorithm a tool that keeps the public informed and holds power to account.

Since these signals are weighted relatively heavily in the algorithms on the front page, a news story with high news value will rise to the top of the page — even if the reader has no particular interest in the topic.

This dynamic ranking allows us to fulfil our journalistic mission while remaining relevant to individual readers. The number of personalised stories users are exposed to is dynamic, so more personalised stories appear on the front page in quieter periods than when several major news events occur simultaneously.

Capturing the nuances of journalism

Binary categorisation of content into “interesting” or “not interesting” buckets for a particular user doesn’t capture the true nuances of personal interests.

For example, whether someone is interested in a particular sport varies depending on the nature of each story, so we must ensure internal categorisation doesn’t prevent readers from discovering content they’re interested in.

One of the most effective methods in our existing rankers is often called collaborative filtering or “what readers like you have also read.” The system helps us avoid binary categories like “sports fan” and “non-sports fan.” It recognises that articles about youth sports, for instance, can resonate with people more interested in family life.

The benefit of this system is the algorithm recognises how topics intersect without understanding the content itself. Instead, it deciphers user reading patterns and bypasses factors like location or demographics, avoiding categorising content into distinct sections (sports, politics, or culture).

After years of experimentation with different approaches to algorithmic ranking and personalisation, we now get the best of both worlds, ensuring everyone sees the most important stories of the day while also helping them find niche topics and discovering the breadth of our coverage through personalisation.

What the results say

Thanks to the blend of human-in-the-loop personalisation and existing rankers, subscriber click-through rates (CTR) on front-page articles increased by about 25% in the past 10 months.

One iteration of the algorithm, which ran for subscribers for 40 days, showed the personalised front page outperforming the non-personalised baseline by more than 8%. This is historic and vastly different from how personalisation is discussed in other contexts, which typically measure percentage lift on individual placements on the front page. This is a lift in overall CTRs on the entire front page, affecting the reach of all content and users daily. 

The effect on gender and age segments is particularly encouraging: CTRs among women aged 30 to 40 were consistently 13% higher for those receiving personalised feeds than the same segment without personalisation. By topic, minority user segments engage with a wider breadth of content in the personalised version, meaning personalisation helps us serve these users better than a feed that suffers from the gender/age bias resulting from optimising for the older male majority. 

This prompted us to test a smaller share of manual placements on the front page. The results confirmed our hypothesis: Switching from fully manual to algorithmic control increased CTRs on the placement by 20%-30%. We now see promising results from manually selecting as few as only the top three articles and letting the algorithm work in sync with newsroom signals to do the rest.

A higher share of personalised placements also exposes users to nearly twice as many article teasers, as measured by unique impressions per user.

Our findings indicate algorithmic ranking and personalisation help us better connect our users to their niche interests rather than optimising for an average inevitably affected by selection bias. When faced with the results of these tests, it is getting harder than ever to find arguments for insisting on a one-size-fits-all answer to our journalistic mission.

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