In theory, the preseason polls should be nothing more than a sideshow, as polling respondents will update their information as games are played throughout the season. The logic goes, by the end of the season, the respondents are evaluating the total body of work for a team based upon careful observations of each team's behavior.
However, the preseason polls represent something very important: the polling respondents' expectations of how teams should perform. In other words, respondent bias. If Craig James ranks USC highly in the preseason, he believes that USC will be good. Hypothetically, if USC and Baylor than provide equal evidence of their quality by the end of the year, but Craig James did not expect Baylor to be good and did not include them in his preseason top 25, he will be predisposed to rank USC higher for no other reason than his preseason faith in Lane Kiffin.
To illustrate the presence of this bias we take three pieces of evidence: 1) the preseason AP poll, 2) the final regular season AP poll, and 3) the final regular season Network Rankings which evaluate the total body of work based on wins and losses (more here, but any evidence based ranking should work). I've collected the data and calculated the Network Rankings for the 2012, 2011, and 2004 seasons. To demonstrate preseason bias, we expect to find that where there is dramatic divergence between the final regular season AP Poll and the Network Rankings, the preseason poll is influencing the outcome.
For example, in 2012, Boise State was ranked 19 in the final regular season AP poll and 40 in the Network Rankings. Because it was included in the preseason top 25, polling respondents ranked them far more highly than they deserved. Alternatively, we find that in the case of Texas A&M, unranked in the preseason, the Network Ranking is 8 and the final regular season AP poll 9. A difference of only 1.
To demonstrate this effect statistically, I regress the final regular season AP poll on the preseason AP poll and the Network Rankings. If the AP poll is simply capturing perceptions of a team's body of work, than the variation in final AP rank should be mostly accounted for by the Network Rankings. However, more likely, the final AP poll is a combination of a polling respondents' perceptions of the actual evidence combined with the respondents' bias of how a team should perform represened by the preseason AP poll.
Indeed, in this (very) simple analysis, the preseason polls do appear to introduce bias. As we would hope, the final regular season poll is contingent upon the evidence, and the Network Rankings are positively and significantly correlated with a team's final AP rank. However, despite controlling for the Network Rankings, the preseason poll is also significant, increasing a team's end of season rank, on average, by 0.42. For those interested in the statistical details, the r-squared is .49, both variables are significant at .01, and in coding the AP polls because they only rank the top 25, you can code the unranked teams as 26 (or code them as 0 and invert the ranking values), or 120, and you get about the same results.
Either the preseason poll represents some information we just can't pick up by evaluating the evidence that Craig James, in his sage wisdom, can, or the polling respondents are biased at the outset and maintain some portion of that bias throughout the season.
So who benefits from this preseason bias? Not the SEC, likely because the SEC usually provides actual evidence that meets even the most lofty expectations. Here's a list of the top 10 offenders in the three seasons of data: