This is the link:
http://www.washingtonmonthly.com/archives/individual/2004_11/005178.phpThe data comes from the Century Foundation, which is a relatively progressive foundation. The data could also be verified from newspapers from those earlier years, if people care to check for the UNWEIGHTED numbers (since the FINAL exit poll data is always weighted to match the election results - just like this time).
Year 1988: Unweighted exit poll Dukakis: 50.3%, Bush: 49.7%. Actual vote deviation 7.7%.
Year 1992: Unweighted exit poll Clinton: 46%, Bush: 33.2%. Actual vote deviation 5.6%.
Year 1996: Unweighted exit poll Clinton: 52.2%, Dole: 37.5%. Actual vote deviation 8.5%.
Year 2000: Unweighted exit poll Gore: 48.5%, Bush: 46.2%. Actual vopte deviation 0.5%.
In each year, the unweighted exit poll results always overstated the Democratic vote, sometimes by MUCH more than 2004.
If people are going to continue to claim that this year the unweighted exit poll proves fraud, then they must also show fraud for every single Presidential election since exit polls started (in 1988). They have to show that such fraud took place 1) at a time when there were no electronic voting machines, 2) in a manner which didn't actually get the Republicans the win twice against Clinton and once against Gore (since Gore DID win the popular vote), and in one of the remaining years in a manner which helped them not at all - since they would have won anyway.
The most logical explanation for all of this is that exit polls tend to favor Democratic voters until they are weighted, and the unweighted data is not helpful in calling a close election.
Again - my usual caveat. I believe there was election fraud - I just don't believe exit poll analysis is in any way helpful in showing that fraud. In fact, it's harmful because it is so easily refuted.