Their Stats Engine does resolve the issues, but I have to laugh at their marketing materials calling it 21st century statistics, because even the new approach is comically behind the times. Looks like the poor schmucks accidentally outsourced the project to some really old-school one trick pony "let's throw some maximum likelihood theory at this" statisticians. I would've hoped even die-hard frequentists would see the value of a multi-armed bandit and the irrelevance of declaring a winner rather than quantifying the effect of each variant and the uncertainty around it.
Cf. the technical paper: http://pages.optimizely.com/rs/optimizely/images/stats_engin...