Hacker News Reader

Doesn't mean it can't be statistically valid, the noise just means you need to run more tests.

That assumes the noise is uncorrelated. It's not hard to imagine scenarios where it's correlated.