Audiobooks.com: Churn Prediction at 95%+ Accuracy in Five Weeks

Audiobooks.com is RBmedia’s flagship digital-audiobook platform, serving libraries, schools, and retail with 250,000+ exclusive titles. The business metric that matters most is Cost-per-Circulation. Every lost subscriber moves CPC the wrong way.

01 The Challenge

Churn is the primary driver of platform economics — but the team was flying blind

Audiobooks.com operates in a market where customers can switch platforms with a tap. Every lost subscriber reduces book checkouts, worsens CPC, and weakens the company’s standing with distributors — particularly libraries and schools, whose government funding tracks checkouts.

Existing churn-prediction work was exploratory. The team wanted a production-ready model they could act on before a customer left — accurate enough to justify targeted outreach, explainable enough that marketing knew why to reach out.

02 The Approach

Evaluation framework first, then model

Provectus opened by reviewing the existing churn work and standing up an evaluation framework — data quality checks, hypothesis validation, reproducible model testing. That gave the team a clean baseline and a structured path to improvements.

The dataset was assembled next: 2.5 million user profiles, five years of listening activity, search and recommendation logs, subscription renewals, and churn events.

03 The Build

Amazon SageMaker, 100 predictive features, explainability built in

Automated pipelines turn raw data into a deployable model at each iteration. Training, evaluation, and scoring run on Amazon SageMaker. Experiment tracking and reproducibility sit on a pipeline-orchestration layer on AWS.

At each iteration, new predictive signals were added and measured against baseline. The final model has 100 features covering subscription behavior, recency of activity, search frequency, and credit usage. Accuracy clears 95% across test scenarios, with several variants reaching 98-99%.

Crucially, the model does not just predict; it explains. For each at-risk user, the marketing team can see which factors drove the prediction, so outreach is targeted rather than generic.

04 The Results

Targeted retention instead of blanket campaigns

95%+

Production churn-prediction accuracy

Across test scenarios

The model shipped five weeks from prototype to production. Marketing can now reach the right customers with the right message at the right time — without contacting users who were never at risk. Reduced attrition tightens CPC and strengthens the company’s position with distributors.

The team owns everything that shipped: the model, the pipelines, the infrastructure. Audiobooks.com runs and refines the model independently and extends it to other ML workloads on the same substrate.

05 What’s Next

ML infrastructure reused beyond churn

The SageMaker substrate and pipeline orchestration carry forward to the next ML use case. Churn was the first workload; the infrastructure Provectus built is the shape for what the team adds next.