Yeah, golden cohorts can work, but they are really hard to pull off, especially for logged out traffic (which is where you'd use most of these patterns anyway). Good luck tracking me over 6-24 months across different devices and locations. And cross-contamination is hard to prevent (for instance, the golden cohort might suffer from global effects like worse content due to loss of power users or even from stumbling across brand-damaging threads like this). It also just adds a lot of product complexity to keep behavior around that long.

That said, they can work. Twitter famously did something like that for their time-based vs algorithmic feed and I think YouTube does it pretty regularly.

The biggest issue, though, is that by the time you get results from any long-term experiments, most of the decision-makers (PMs, EMs, etc) have probably moved on away having taken credit for the short-term wins they delivered.