there's always sample bias - estimating and adjusting for it is the tricky part of course. The (statistical) "correction" for it is to insure the polling crosses all boundaries - and that's probably impossible - I know for example I have never particpated in a telephone poll - the interruption pisses me off for one thing - and I know many others like me in that regard. I have been called - and even answered some questions - but it rapidly became obvious to me that the wording of teh questions was tilted in some manner - so I declined to futher play that game. Being online certainly targets a certain audience - however the model discourages excessive influence (unless you've got deep pockets), and yet incurs such a small opportunity cost that anyone who
desired can easily participate. My untested hypothesis is that anyone who makes an effort (that is minimal but more major than simply clicking) to participate when there is an opportunity cost will likely actually go vote for real - those are whom I actually want to sample.
The only true test will be to compare after the election.