The other part is we found that people demoing the software are less likely to follow the instructions (since they're not getting paid) and we worried about having a lot of poor tracks that would look bad. Our testers have been amazingly good at following instructions :).
1) Our processing can't be done realtime (i.e. it takes more than 1 second per second of video), at least for now, so there would need to be a lag that can get significant if they user has a slow enough computer.
2) We currently don't require any software to be installed client-side (we use Flash + JS), which we consider a big perk.
Looks like you have a bunch of tough optimisation ahead, much good luck with that.
Very nice idea by the way, eye tracking used to be the domain of specialised setups using cameras behind half-transparent mirrors, it's quite impressive that you've gotten this far with just a cheap webcam as an input device.
I could imagine that using CUDA-enabled kernels might be of help for processing images.