Except, apparently, they aren't actually that good at _that_. If an A/A test to not yield 100% chance of 18% uplift, what gives you any degree of certainty that other tests won't have equally skewed results?