My guess is that “tens of thousands” is the rough SELECT count(*) from their systems. And I’d bet most of these experiments are old Boolean feature flags that were essentially acceptance tests, the final hurdle before global launch. For example, seeing that your new Gmail POP3 settings workflow drives engagement down 50% is a pretty clear indication you’ve got a bug that wasn’t caught in development. The interesting tests would be buried between such mundanities.

This kind of information would also be highly proprietary so I doubt such a list would see the light of day.