It's going to be tricky to get meaningful quantitative data comparing a live site and a prototype (or even between two sites). Especially if they're significantly different from each other. Even if the measured times and success rates are significant, how do you determine what factors caused the differences?
A Q&A survey has all the drawbacks that surveys have:
- They typically rely on users recalling what they did and why, and that's unreliable.
- Survey takers sometimes skew their answers toward their own points of view and what they think the moderator wants.
I'd suggest staying away from surveys and timed tasks.
Rather than timing users as they perform tasks, you can quantify the number of actions needed to complete common tasks. With Keystroke-Level Modeling you'll go through the tasks and note exactly what clicks and keystrokes they require to complete, and the time associated with each. You'll end up with quantities of actions and surprisingly accurate times for task completion. These you can compare. (This is done without users; you'll probably do it yourself.)
KLM is an old method, but I was always impressed at its accuracy at modeling expert behavior.