JAMES - Your Digital Butler Driving Subscription Growth Using AI
Media associated with this campaign
Overview of this campaign
“JAMES, your digital butler” is an AI technology developed collaboratively by The Times and Twipe. During a 1-year project, a team of 20 people over 3 countries worked on building JAMES.
JAMES served over 100,000 subscribers of The Times with individualised newsletters compiled from the content of a daily edition. 18 optimisation algorithms were developed and tested: time optimisation, content recommendation and format optimisation, including innovative bandit-driven next best action communications. The Times observed 49% churn reduction among subscribers using JAMES. Furthermore the technology is made available to the wider industry through the Twipe digital publishing platform.
Through initiating this research, we wanted to leverage the richness of the 1 billion data points generated on our domain everyday in order to create more meaningful experiences. We wanted to focus on discoverability and convenience of our edition content reach and aim to close the perceived value gap of low to medium engaged subscribers. Additionally we sought to explore implementing AI and ML in a large matrix-style company and explore how other newspaper publishers might benefit from such technologies.
JAMES stands for Journey Automated Messaging for higher Engagement through Self-Learning and has the objective to accelerate our net subscribers growth by individualising the way we distribute the content of the editions to readers.
JAMES is "a digital butler" and the metaphor is not coincidental. Like a butler, JAMES learns and knows the preferences of the user and discreetly serves the right services to them as individuals, rather than at the segment level, which was the most advanced operationally feasible practice for customer marketing. Powered by self-learning algorithms JAMES analyses and learns how readers consume our edition and provide them with individualised newsletters which are then continuously optimised to increase engagement.
The individualisation algorithms behind JAMES are based on four dimensions of optimisation.
- Time: the best time to send an email.
- Content: which content triggers reading.
- Format: layout and copy.
- Frequency: the number of emails sent.
Throughout the project time frame, six different propositions were tested, combining different models of calculating time and selecting content, and have different formats and frequency of sending.
While personalisation and individualised newsletters have been popular in other industries for a while now (like e-commerce, where the objects to personalise for are static and known), JAMES is groundbreaking for bringing personalisation to news articles which, through the dynamic nature of news being created every day, is much more challenging to accomplish.
Results for this campaign
Throughout the 12 months of the project more than 14 million emails have been sent to the cohort of readers included in the various experiments.
This has resulted in significant impact with more than 70% of the exposed readers interacting with JAMES and a 49% decrease in customer cancellation rate on the selected cohorts.The impact has been highest on customers with low to medium engagement.
Each day, during a period of 10 months, 60,000 subscribers who didn’t receive a newsletter at that time were exposed to different variations of the JAMES Daily Briefing Proposition.Every week, the cumulative number of people who interacted with JAMES by clicking or opening increased to more than 70% at the end of the project. Furthermore the rapid stabilisation of the amount of opt-outs, ending at approximately 15%, was also very encouraging for the team.
Readers who click through on the JAMES Daily Briefing Proposition have increased their engagement with The Times significantly. Moreover, the more they interact and engage with JAMES, also the larger the increase in their engagement with The Times.
The JAMES Daily Briefing Proposition works better for low to middle engaged subscribers.
Knowing that increasing engagement amongst low engaged customers is key concern across the industry, it was a great result to see an increased click-through rate for this group compared to highly engaged subscribers.
In terms of newsletter composition, the best performing combination is a morning email with a mix of popular and recommended content.
Layouts that limit the number of images and scrolling time in order to make email content easier to consume or formats that have a larger number of articles worked particularly well. One format displayed as many as 20 articles where a very high click through rate was observed.
In terms of content, we found that the closer an article topic matches to a reader’s personal interests, based on past reading behavior, the higher the click rate, but also that the more popular an article is - given that it has not been read yet - the higher the probability of receiving a click.
The cases when JAMES could find a combination of both factors in one: namely serving an article which match personal interests and is also popular to an individual could generate up to 8 times higher click rates.
As a result we developed a hybrid model which combines the results of popularity and content personalisation.