Content People Actually Watch
a mid-market streaming platform
viewer engagement
reduction in monthly churn
more of the catalog watched
Viewers browsed for minutes, watched one title, and left. The existing recommendation system pushed the same popular titles to everyone, so most of the catalog never surfaced and licensing spend on that content was wasted. Churn climbed as subscribers concluded there was nothing to watch, when in reality there was plenty they simply never saw.
How we approached it
Analyzed viewing sessions to identify where browsing turned into abandonment and which catalog segments never surfaced.
Built two-tower embedding models that match viewer taste profiles to content, updating in real time as each session unfolds.
Solved cold start for new and niche titles with content embeddings generated from scripts, metadata, and visual features rather than watch history alone.
Stood up an experimentation pipeline so every ranking change ships behind an A/B test with engagement and retention as the deciding metrics.
“People stopped scrolling and started watching. Titles we thought were dead weight found their audience, which changed how we think about what to license next.”
New titles get content embeddings built from scripts, metadata, and visual features, so they can be matched to viewer tastes from day one instead of waiting for viewing data to accumulate.
Engagement combines watch time per session and titles completed, measured through A/B tests against the previous recommendation system rather than before-and-after comparisons.
Yes. Rankings update in real time from what a viewer browses, samples, and finishes during the session, so the homepage reflects current intent rather than last month's habits.
Want results like these?
Book a free consultation and we'll map the highest-leverage place to start for your team.