On the other hand, if you can eventually measure long-term value and you can keep track of users over a long enough cycle, then you can eventually run an A/B test on long term value. This depends on accurately tracking users, having a quantifiable way to calculate long-term value, and eliminating other sources of error, but it should be absolutely doable.

You can also measure how well your proxy stat does correspond to long-term value and adjust your expectations accordingly.