Netflix was an early adopter of personalisation.

What began as a relatively primitive recommendation system rolled out in 2000 evolved into profiles that allowed individuals in the same household to create profiles and select their preferences.

By 2108, it began personalising visuals to make selections even more appealing, Sonali Verma, lead of the INMA Generative AI Initiative, said.

“We’ve seen parallel changes in the news business as AI has become pervasive,” Verma said, noting that companies began working on recommendation engines about 15 years ago and “played with” things like collaborative filtering, geolocation, dynamic paywalls and offers, and more.

Initially, such experimentation was costly, Verma said: “You needed a serious data science and engineering team to build this, to make any of these systems work, and then you had to maintain them.”

But, with the explosion of AI, all that has changed. During the recent INMA Generative AI Master Class, media leaders from Politico, Aftonbladet, Keskisuomalainen Oyj, The Hindu, and The Globe and Mail shared how their companies are leveraging AI to personalise their products.

Politico transforms policy research

When the team at Politico sat down to brainstorm the future of AI in its organisation, it turned its attention to its digital Politico Pro platform.

The subscription-driven platform attracts companies, organisations, and government agencies who use it to “be better at their job,” Caitlin Bugas, associate director of product, said.

“There’s things like policy trackers for legislation and regulation,” Bugas explained. “We have an editorial team that’s dedicated to providing real-time news and intelligence, and we also have a suite of data products available there.”

The massive amount of information it contains made Politico Pro “an excellent playground for us to test out different AI concepts and sort of determine what would be appropriate to scale across our broader business.”

Politico’s first AI experiment came in the form of bill summaries. Given that U.S. legislation often runs 3,500 words or more — and sometimes over 15,000 — summarising these dense texts was an obvious opportunity.

The team launched AI-generated summaries directly within Politico Pro’s legislative analysis tool, clearly labeled as AI and in beta. It also included a feedback form to monitor user response.

The results were eye-opening. The feedback proved that with the right transparency and utility, AI could be welcomed by Politico’s audience.

“We received both a high volume of qualitative usage and quantitative feedback,” Bugas said. “It validated that this was a super big interest for folks.”

Aftonbladet’s election chatbot

One of the biggest AI success stories for Aftonbladet in Sweden is its Election Buddy chatbot, which was originally built for the E.U. elections. It was so popular that it was updated and used for the U.S. elections, Martin Schori, director of AI and innovation, said.

The chatbot was trained on thousands of questions and answers based on the company’s journalism, and, in addition to answering questions, it provided a fun and entertaining way for readers to educate themselves.

The results were “pretty crazy,” with the bot answering more than 600,000 questions in just a few weeks. The success of the chatbot led Aftonbladet to broaden the scope and launch a chatbot covering all of the company’s content.

In two years of working with GenAI, the most important lesson Aftonbladet has learned is also what may seem the most obvious, Schori said: “The quality is everything.”

People are sceptical about using AI, but if you provide them with the proper tools, they will use them. That puts the responsibility for developing and providing quality products and tools back on news media companies.

“I mean, we are responsible for everything we publish,” Schori said. “So if we put out bad products on the market, that’s our fault and not AI’s fault.”

AI to personalise newsletters at Keskisuomalainen Oyj

Setting out on a mission to better serve a younger audience, Keskisuomalainen Oyj in Finland began using AI to personalise their newsletters.

They saw an opportunity to optimise this channel of delivering news considering the newsletters they had been sending out weren’t functioning very well.

Content recommendations are based on a person’s behaviour and the trending status of articles. They also take into consideration if a user has already read certain articles and only include content that has been published within the past 48 hours.

“We are a regional and local publisher and people are really looking for the local news,” Johannes Niemeläinen, digital insights team lead, said. “If we feel the newsletter with news agency content or opinion pieces or comics are not really read that much by our subscribers, then it will really not function that well.”

When they began collecting data on their users, Keskisuomalainen Oyj got a bit of a surprise. Their goal of attracting younger audiences through this new channel was not the reality.

Out of 113,000 subscribers, the mean age of active newsletter readers was 66 years old. They thought this channel would be appealing to working age people who could read a newsletter on a lunch break but that wasn’t what was happening. They also saw that younger audiences were more inclined to opt out of the newsletters.

The actual silver lining here? The older audience was really engaged.

“More than half of the people getting these or receiving these e-mails open it once a week, which I find quite a good result,” Niemeläinen said. “So there’s this really engaged majority and a group of super active readers reading the personalised e-mails.”

The Hindu’s personalised push notifications

“Personalisation is no longer just good to have,” Pradeep Kumar R, senior product manager at The Hindu in India, said. “Today, it’s a necessity. Users expect content that is tailored to their interests, and without that, they quickly disengage.”

Before implementing a new “extreme personalisation” model to power its push notifications, The Hindu was using an “old school system” of tagging content and segmenting users, recommending stories based on broad categories.

There were several challenges with this system, including an outdated taxonomy and limited user interaction points. The latter was leading to inconsistency with recommendations, which resulted in less engagement.

Kumar said there are three recommendation strategies:

1. Collaborative filtering: Clustering users with similar reading behaviours and suggest content based on shared preferences.
2. Behavioural filtering: Matching users with articles by measuring cosine similarity between content embeddings and a user’s interest profile.
3. Hybrid model: Balancing trending news, editorially curated content, and user-specific recommendations.

“The shift from basic rule-based recommendations to AI-driven personalisation has been a game-changer for us,” he said, “both in terms of engagement and retention.”

Before, push notifications were manually triggered when new content was published — which is not necessarily when users are interested in receiving new content. Now, notifications are triggered based on when and how the user interacts with the content, the click-through rate has increased eightfold.

The Globe and Mail personalises its homepage

At The Globe and Mail in Canada, an initiative to personalise its homepage began with recognising the inherent challenges in digital news consumption, especially the “friction” users encounter when trying to find resonant content.

Mike Pletch, managing director of product and UX, described this friction as the difficulty “for people to find stories of interest,” which often leads readers to “bounce around the app and Web site” in search of relevant content. This not only consumes time but also diminishes the user experience.

The Globe and Mail has now refined its homepage personalisation by implementing a comprehensive process that begins with creating a candidate pool of articles. This pool includes both newly published and evergreen content. Each article is then meticulously scored and ranked based on engagement, business value, and personal relevance to individual users.

Finally, the remaining articles undergo final assembly where they are matched to specific zones on the homepage based on editorial guidelines and rules. This step ensures that the homepage remains organised and easy to navigate while still being dynamically personalised according to the structured placement of content.

Through this methodical and strategic approach, The Globe and Mail strives to balance personal relevance with editorial importance and business value. This results in a homepage that is engaging, informative, and aligned with user interests and behaviors, effectively merging data-driven insights with editorial acumen to enhance the overall reader experience.

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