Diagnose and Fix Position Bias in a Bandit-Ranked Carousel

A content app uses a contextual bandit to rank 5 content items in a home-screen carousel. Six weeks after launch, an analyst notices that whatever item the bandit puts in slot 1 has, on average, a much higher observed click-through rate than the same items get when placed in slot 3 — even comparing the same item across different impressions.

1. Explain the mechanism by which this creates a self-reinforcing problem for the bandit's reward estimates specifically (not just a general observation about position bias).
2. Propose two concrete fixes, drawing directly on techniques named in this subject, and explain how each interrupts the loop.
3. Explain precisely why the bandit's own UCB or Thompson-sampling exploration mechanism did not already prevent this.