01
Personalized recommendations for the SWR website
AI recommendation system in production · Product Owner · SWR/ARD in collaboration with ZDF
Starting point
Curated by hand or from a legacy system. Neither scaled.
The recommendations below every article play a big part in whether readers stay on the site or leave. Manual curation tied up editorial capacity that was needed elsewhere.
Approach · The feedback loop
01
Model shows two recommendation variants
02
Editors click the better one
03
Feedback as a training signal
04
Model improves
The evaluation tool
Radically simple.
The real problem was not the technology but the editors' time. Two variants side by side, one click on the better one, done. A few seconds per rating, no onboarding needed.
Two recommendation variants · one click on the better one
Live on SWR.de
Result
Measurably better than the legacy system.
After go-live, we compared views and clicks on the recommendations with the historical numbers from the legacy system. The new system ran in production permanently.
Legacy system
New system
Clicks on recommendations, relative · no axis values
What I take away
AI projects rarely fail because of the technology.
They fail because the people whose knowledge you need have no time. If you make giving feedback as easy as possible, you actually get it.