Yesterday, I wrote about active (or declared) personalisation. The other option is passive (or inferred) personalisation — passive (on the users’ part), where we use technology to make recommendations based on past reads, user behaviour, demographics, and other data-led signals.
Karl Oskar Teien, director of product at Schibsted, shared the graphic below with me. It shows some of the tags we can use and considers the balancing of editorial signals and user preferences.
The South China Morning Post uses these four user variables to base their personalisation on:
Arguably the best company in the world at passive personalisation is the video app TikTok. At no point in the process do you select what you want to see. They take signals based on behaviours to determine what you do and don’t want to see.
Nils Schimmelmann, a senior engineering manager at Yahoo USA, took me through the basics of how passive personalisation works:
This is a simplistic view of what can easily become exceptionally complex. It helped me to understand the principles of personalisation but then led to many more questions:
Do we prioritise solely on what the user wants? Should we be showing a diversity of information to avoid filter bubbles? How do we build in business needs? Are we optimising different goals for different readers, such as articles with high conversion for non-subscribers, building a daily habit for new subscribers, or advertising yield for those likely never to subscribe? Should we be looking at the sequencing of content, such as to give some lighter content after hard news?
Once you start thinking about all these questions, there are many rabbit holes to go down. If you are at that point, I highly suggest following my colleague Ariane Bernard’s work in the INMA Smart Data Initiative. She has extensive knowledge on these systems and is going deep into the subject this summer.
Yahoo has built its own algorithm in house, which, in parts, uses open source software. But if you are starting from scratch, there are some out-of-the-box tools you can experiment, with such as Google Rec AI.
Another question that has come up is around how often the feed is refreshed or how far to go back in time. A benefit of using an algorithm is being able to revisit evergreen or older content that may not get resurfaced otherwise. Mediahuis Netherlands is one company that was surprised by the range of content its experiments brought up. In the results on initial tests, the algorithm pulled from 564 articles whereas the manually range was 267. The team also saw a change in the content mix: less news, more opinion, more video, more female-focused content, and more sports.
Passive personalisation can also be used to enhance active personalisation. For example, when asking which topics a user wants to follow, you can add or replace with suggested topics based on the information you have about them.
If you’d like to subscribe to my bi-weekly newsletter, INMA members can do so here.