Comment by tjbai

I agree that there's an exploration-exploitation tradeoff, but for what you specifically suggest wouldn't you presumably just normalize by sample size? You wouldn't allocate based off total conversions, but rather a percentage.