I know that this every productivity measurement is very fuzzy and can be cheated (#), but here's how we did it:

We were simply looking at the number of features delivered to the customer in a certain amount of time, number of defects found in production and number of critical defects found immediately after roll-out (a.k.a. hot-fixes). We adjusted for holidays, vacations and sick leaves.

Those numbers were relatively stable until we changed the team, then dropped, and it took quite a while to reliably get to higher levels than before.

(#) I don't think the team cheated, because if they had, they could have easily cheated the numbers in a way that they'd steadily rise.