The paradox of perfect process
summary
A global ML recommendation engine plus AI image recognition ships from a separate POC into the main app and reaches 20 million daily users during a 10-month global rollout, showing 35% conversion uplift on surface metrics while the purchase rate stays flat. Root causes include payment-step data that captures only a fraction of purchases, poor recency weighting, and a feedback loop where confusing recommendations amplify irrelevant clicks. Fixes: treat low-history users as new, add market-popular lists, human curation and stricter signals, and prioritize purchases over clicks.
AI-written summary · about
