I found it very interesting that MAB-90 was outperforming RAND on both convergence speed and average conversion rate in the first experiment. This seems a very interesting result, albeit not the one you might be looking for in the article. I find it rather disturbing that this dataset was removed in later experiments. It actually seems to provide a very valuable insight namely: Multi-Armed Bandit strategies might be better than A/B testing but the exploration factor should be far higher because of the need for rapid convergence to significance.