A news publishing company with +100M pageviews per month asked us to help it increase its reader engagement. Though it was spending significant effort to lift engagement, it was only achieving meager results. The company was relying on old-fashioned manual analysis of engagement results and making real-time editorial choices, which essentially meant that it was “guessing” the optimal article order and applied it uniformly to all readers.
We ran an optimization campaign to showcase the strength of the newly introduced Rumble Px personalization platform to personalize the content being served and automatically adjust the item order.
To begin the optimization campaign, the publisher needed to select the key performance indicator (KPI) to be optimized. The publisher selected Item Views with the goal of increasing the total number of articles read.
Rumble Px in Action
For this specific optimization campaign the publisher selected ItemReflow™, a Rumble Px Tool that controls the order of articles served in each channel, thus offering a personalized item order and increasing user engagement.
Additionally, the publisher selected auto-segmentation, which meant that the Rumble Px platform would automatically determine the optimal segmentation mix, down to a micro-segment of an individual user. In this case, we used a few weeks of curated data on the customer base that contained engagement information of a user with a specific article in a specific article position.
The Rumble Px Platform then ran a regression model based on the ranking of an article (normalized by removing the effect of the article position in the channel). Using the regression model, we were able to predict an article’s performance and we used this prediction to reorder the articles in a channel by their predicted performance.
In applying Rumble Px for the Media Company in a production environment, we applied the optimization to 90% of the users and left the remaining 10% of the users with the default order to serve as a control group. The control group not only served as a performance benchmark, but also allowed us to detect if the optimization group had a specific instance of lower performance, we would be able to react by retraining the model based on that data.
The lift in engagement was almost immediate. The results at the end of the first week showed an average increase of 26% in articles read, with peak engagement increasing +30%.
The graph above shows that the group of users that was shown an optimized item order registered a total item count that was higher by up to 30% as compared to the control group, and registered an average increase in engagement by 26% after the first week.