Statistics you can defend
Bayesian results
The probability each variant is best, and what it would cost to be wrong. Computed for every experiment, alongside the p-value rather than instead of it.
The Bayesian half models each variant conversion rate as a distribution rather than a point, starting from a uniform prior that assumes nothing, and updates it with what has actually been observed. From those distributions it reports the probability each variant is the best of those tested, the expected cost of choosing wrongly, and a credible interval per variant.
It is computed for every experiment, not as an alternative mode you have to opt into. Both halves are on the same screen because they answer different questions and people need both: the probability is easier to act on, the p-value is easier to defend.
The probabilities come from sampling the posteriors many times over and counting how often each variant comes out ahead. One detail matters more than it looks: the sampling is seeded from the data itself, so the same counts always give the same answer. Without that, refreshing the page would move the probability slightly every time, and a number that changes when nothing has changed is a number nobody trusts.
Like the frequentist half, this describes a converted or did-not-convert process, so it is computed for conversions. Revenue per session is not that kind of quantity, and the console reports it as absent rather than filling the column in.