If you start cheating the metrics, or optimizing a lot towards them, it becomes counter-productive when they change. As such, the most efficient way forward would be to work without trying to optimize for a temporary metric. On the flip side, it would be troublesome to convince people to different, complex forms each time.
What first gave me the idea was the concept of "lubricating" headers (submitting some with random values) for future http protocols, to combat "ossification", where middle boxes start to meddle with them and become obsolete when they don't recognize the new fields, instead of transmitting them.