From a friend in the campaign consulting world, Matt Braynard. He’s not one of the world’s big Zogby’s, but his accurate polling methods made me look much smarter than I am when I called certain 2006 races:

I guess you know why / how all those polls were wildly inaccurate, especially on the D side. Every one of them uses a “likely-voter” model based on self-identification. Every one of them. The celebrity-intensity factor for Barack Obama was huge. People who were registered through motor voter, people who registered and hadn’t voted in years, people who weren’t registered at all and didn’t realize they needed to register to vote — they were far more likely to tell a guy on the phone they were absolutely voting than they were to actually find out where they needed to vote and then actually show up and vote. Zogby was most wrong (16 points off) because his operators push people to give an answer when they are really undecided. Sometimes this works, but when there is a huge celebrity factor, someone who is otherwise not inclined to pick a candidate gives the answer they think they are supposed to.