There are a number of reasons why legacy media faces fierce rivals in digitally native media – such as Netflix, Spotify, and Stitch. However, a critical competitive differentiator could be the fact that many of these brands are committed to delivering accurate, actionable recommendation systems.
Recommendation engines aren’t just about recommending content to your readers, according to Michael Schrage, a research fellow at MIT Sloan School’s Initiative on the Digital Economy and author of Recommendation Engines. They are platforms that facilitate online interaction. When designed well, they help organizations rethink how they can get more value from their data. They also offer more value to their customers.
“The biggest mistake media companies make is they think they are in content business. But they are the data business – and their data management sucks,” states Schrage. “While they won’t be able to rival Netflix, they still can engage with their communities so much better. It’s not just about how to tell a better story, it about how we can learn from people who read our stories and engage with them.”
Recommendation engines, or recommenders, are nothing new. Amazon founder Jeff Bezos spoke about the value of recommenders back in 1998. He said that they offer an “opportunity to develop very deep relationships with customers.”
The good news is that, while they were once a feat of extraordinary technological effort, machine learning is changing that. In fact, it’s the ease with which they can be built, applied and utilized that makes them the most exciting and valuable applications of machine learning right now.
“The big difference now is that anyone can build a recommendation engine,” states Schrage. “They are cheaper, faster and better than ever. Plus, the underlying algorithms have moved away from simple correlations to much deeper and richer analytics, around diversity and relevance. As a result, the levels of sensitivity and specificity of recommenders have improved. And they keep learning how to get better.”
Virtuous cycle of engagement
That is the key to well-designed recommenders. The more people use them, the more valuable they become. And the more valuable they become, the more people use them, creating a virtuous cycle of data with audiences. However, according to Schrage, legacy media often miss this vital opportunity to improve both customer insights and engagement.
“These organizations treat recommendation engines as a sales tool. But they need to understand how to use their data and turn it into virtuous cycle,” says Schrage. “And the low-hanging fruit is advice and recommendations.”
In other words, recommendation engines are more than a UX differentiator. When effective, they enable organizations to build customer lifetime value (CLV). The former is transactional, while the latter embraces a longer-term view.
“It’s the difference between ‘always closing’, and a salesperson who wants to forge a relationship,” says Schrage. “The question is: Do you treat customers as ‘one offs? Or do you seek to use data-driven advice to get people to come back and explore, while also building brand credibility?”
Reader regularity is a key factor in preventing churn. The industry has long been concerned about unengaged subscribers. Recent research reinforced this concern when it revealed nearly half of digital subscribers to local news outlets are ‘zombie’ readers who visit a site less than once a month. These customers are far most likely to cancel their subscriptions than are frequent visitors.
The study also found that personalized, local content is a key factor in retaining subscribers – which is where recommenders play a big part. One great example of this is with the Dutch newspaper, NRC. After personalizing its newsletter, using an AI tool, 22% more readers demonstrated habitual reading behavior. However, there is a lot more to recommendations than personalization.
While recommendations can be personalized, it’s not the only – or the best – option. Personalized recommenders require large amounts of data on users, which a system won’t have with new visitors, or those that don’t sign up. This can cause what is commonly coined ‘cold start’ problem. This occurs when a recommender system cannot draw inferences, due to a lack of information. Another problem with personalized recommenders is the potential for bias, which can be a particular problem for news sites.
This was a concern for the non-partisan news organization POLITICO Europe, when it launched Pro Intelligence in 2019. Pro started as a pure news service. However, it bought the start-up Statehill in 2018, so that it could build out a one-stop shop for policymakers. The platform, which accounted for 60% of POLITICO’S income in 2020, now offers news, analysis and relevant, external information within a single dashboard.
Every piece of content on the Pro platform features recommendations, from legislation to court cases, and press releases to news from the editorial team. However, the algorithm doesn’t require user input to customize the experience. Instead, it employs an auto-tagger, which uses key words, dictionary mapping and natural language processing. [Editor’s note: An in-depth look at this is available to DCN members.]
“We don’t have the engineering capabilities in our organization to dedicate to a personalized algorithm. And we’d probably be too afraid to introduce it, in case of bias,” said Karl Laurentius Roos, Director Technology Development. “But our auto-taggers mean we can surface news and data automatically, which has a real impact on the end user’s productivity and understanding.”
Pushing news, via email or mobile notifications, remains POLITICO Europe’s most effective way to integrate their services into their users’ day. However, Roos and his team are currently building products that will pull the user into their ecosystem.
“If we can enable user access to pull more content themselves, we’re increasing the product utility and value captured each time a user accesses our content,” said Roos. “For us this translates into connecting news with data and wrapping it seamlessly across our desktop and mobile experiences.”
According to Schrage, other media brands doing a good job with their recommendation systems include the Wall Street Journal, Dow Jones, Bloomberg and Pocket. However, not everyone agrees that recommenders are beneficial to business. Slate built Myslate in 2015, which worked on personal recommendations, based on users telling them what they wanted, rather than machine learning. After an initial experiment, the online magazine decided that keeping the infrastructure running wasn’t worth the amount of reader minutes it generated.
“For the recommendations to work we needed to ask them to sign up, which added friction to the experience,” explains Greg Lavellee, VP, Technology at Slate. “But I’m not convinced that recommendation systems are worth it. People tend to personalize their own experience. They know where they want to go on the site. They find the content they want – you don’t have to shove it in their face.”
Try telling that to the likes of Jeff Bezos, Reid Hastings, and Katrina Lake. Nonetheless, media brands may be concerned they don’t have the capabilities, budget or data to build a successful recommendation system. However, Schrage says the issues aren’t necessarily technological or financial, they are organizational and cultural. Recommendation systems don’t just require an IT upgrade, they require a rethink of business fundamentals.
“Legacy media have had the sh*t kicked out of them by companies that literally didn’t exist 10 or 15 years ago,” he states. “It’s because the tech giants have committed to using data to learn. They take good design, good data and good recommendations seriously.
“This is what other media brands need to emulate in order to bridge the digital divide.” In fact, he says that media companies need to learn how to leverage customer’s behavioral data to drive value and grow business. “Algorithmic plagiarism is essential to future health of legacy media.”