Wow, that's a super in-depth answer, thanks a lot! I must admit, I was considering MCMC as an option of last resort (I'm the only one of my team with background in MCMC) and had hoped that someone came up with a miraculous trick for a similar scenario. I'll give this question two more days, and if nothing better appears, I'll accept the answer - I think it's a very nice example of bayesian approach.

Session is the randomization unit in this case. Both conversion rate and average transaction value are important, but when impact on them has an opposite sign, revenue per session must be looked at. And building confidence intervals aground that metric is non-trivial, hence the question here.

Adding a spike ad zero increases the variance a lot and makes experiment insensitive.