Two experiments carried by mid-sized publishers in Europe and in Asia provided evidence of promising performance for paywalls targeting users based on propensity modelling.

The results of the experiments were shared exclusively with INMA by Deep.BI, an analytics platform provider, and presented at the INMA-supported GNI Subscriptions Lab last Friday.

• During the AB tests, carried out between January and May 2020, the visitors of both Web sites were split into two groups in the proportion of 70 to 30.
• Group A was a control group treated with an existing paywall solution: a freemium model in case of a European publisher and a metered model in case of an Asian publisher.
• Group B was a test group treated with a paywall targeted based on a propensity to purchase model trained with the data of respective publishers.
• The European publisher with the freemium model observed a conversion rate lift of 166% and the Asian publisher with the meter — 77% lift. 

Based on my 2020 study of the most popular paywalls, only 5% of news sites in 33 countries were hybrids that combined different models such as a freemium and a meter, or had a dynamic paywall targeting readers based on behaviours and other attributes.

The freemium model is by far the most popular among news brands — 47% of sites tested in 33 countries had some content free to all users and some premium content accessible only to subscribers. The meter model, in which readers can read a limited number of stories for free, is followed by 12% of brands. 

Targeted paywalls stop readers selected with advanced analytics techniques, such as data mining, statistics, modelling, and machine learning.

• In a nutshell, data scientists collect the historical or live data from different sources: for example, online traffic data, content data and customer data.
• After cleaning the data, they analyse it algorithmically to identify patterns, such as the behaviours of readers who subscribed to the site in the past.
• Then they can deploy statistical models to predict future outcomes based on similar patterns, for example they can calculate the likelihood for a new reader to purchase the subscription.
• The segment of prospect subscribers can be then fed to the paywall or e-mail management software to target the offers.
• According to Michal Ciesielczyk, head of AI engineering, Deep.BI uses an ensemble of machine learning algorithms such as Gradient Boosting Machine and Random Forest in its predictive analytics.

Being a registered user to a news site is found to be universally the most predictive feature indicated by five propensity to buy models that Deep.BI deployed at mid-sized news publishers, local and national, in Europe and Asia.

• The ranking of the predictive features is followed by interactions with the paywall such as clicks on an offer or the frequency of being stopped.
• Then come temporal variables such as a loyalty score (time since the first visit ever), frequency of visits, habit score (regularity of visits), time spent on the site, as well as a short- and long-term changes in the engagement score called RFV (measures recency, frequency of visits and volume of articles read).
• Other features found predictive are: whether a reader signed up for a newsletter, the number of shares on social media, and the number of images clicked and videos played.

According to Ciesielczyk, the biggest challenge in building successful propensity models is the amount of behavioural data that can be attributed to a single reader. It is easy when the reader is registered and logged in, as her every action can be tracked. It is much harder when she is anonymous and tracked only with cookies.

Deep.BI found a minimum of 1,000 transactions and six weeks of traffic data is required to train its propensity models. 

Share your case study with INMA peers: Have you deployed predictive modelling? What have you learnt so far? E-mail me at: grzegorz.piechota@inma.org 

Banner image courtesy of Mediamodifier and Free-Photos from Pixabay.

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