Even if that were not the case, exit poll samples are numbered in the low tens of thousands (I believe CNN had n~20,000 for the 2008 presidential election), which would yield a margin of error not dramatically different from this study. Despite having far fewer responses, once you get into numbers this high the confidence is quite large: n=20,000 has a 99% margin of error (for a 53% statistic) of 0.9% and n=120,000 has a 99% margin of error of 0.4%. Naturally, that assumes a perfect sampling distribution which neither the internet survey nor the election polls were fortunate enough to have.

demographic election data comes from exit polling, not the ballots themselves.