First off, you're right: any metric can be gamed. Even so, Google is already using performance to factor into its algorithm, so some level of reasonable accuracy must have already been agreed upon.

Ensuring a certain level of performance is not a very easy problem to solve. But it's doable.

Myself and a few others sat down with the AMP team awhile back and chatted about what that would look like and there were definitely a few different layers that would be needed to get it to work.

The first is already in progress. There's a spec for something called "Feature Policy" (https://wicg.github.io/feature-policy/). Feature Policy would let you tell the browser you don't need certain features which would in turn let browsers take shortcuts (so to speak).

For example, you could declare that you will not be allowing document.write. Enough features being declared could serve as a way for Google (and others) to say "Ok, this is going to be reasonably fast."

There's more needed of course: we had discussions about a standard related to declaring a limit on asset sizes, etc. But it's a start. And while it will always be a little fuzzy, the same is true of AMP. It's completely possible to build a slow AMP page.

The best Google can do is say "if they're doing X, then the odds are good that it's performant".