Hmm.. Here were my thoughts on the matter of voluntary opinion polling.
1) The site validates the user. (Telephone, DOB, last 4 of social, an email address, and a PIN.)
2) The site allows ONLY ONE submission per election cycle from that user. It tells you this-- VERY PROMINENTLY.
3) The site allows a user to notify staff that their one shot has been compromised by an electioneering robot, which gets staff involved.
4) When actually reviewing the submissions, random sampling is used*, along with a real human interpreter**.
* If given a sufficiently large dataset to work with, a randomized sample from that dataset should be roughly approximate of the rest of the dataset. By this, I mean sampling every Nth entry, and evaluating them, would reasonably approximate the results of having every sample evaluated. The difference is how much confidence you have afterward.
** Because we are dealing with human-produced, natural language submissions, we really need a human interpreter to condense the submission into a form the statisticians can use. The costs in time and labor to do this are a significant factor in why we do the random sampling. Human interpreters are given guidelines on how to treat a submission, so that it can be properly tracked for such things as "Degree of partisanship", "Seriousness", "Originality (EG, how likely this submission was actually generated by a real human, and not a robot)", and a generalized assessment of the content of the submission itself. (Support for candidate? Specific grievances? Specific concern? etc)
This, coupled with traditional polling data, would give a deeper picture of the electorate's views and reasoning behind their selections, and their selection process. It would also give insight into the percentage of people responding to the poll that are easily manipulated by memetic PR, and even which kinds.