For those confused by why this poll has such a large divergence between "Registered" and "Likely" voters (especially when the pool difference in only 900 RVs against 791 LVs), this is because the USAT/Gallup polls use a very unusual method of determining likely voters. Rather than ask whether people are likely to vote, Gallup (but not Gallup tracking) simply reweights their poll based on how likely they think people are to vote. This could be more accurate in the sense that it avoids the small number of people who may say that they are certain to vote but will actually stay home, but it is overall a largely untested--and unconventional--method.
I always thought a lot of LV screens were based on past voters?
I like "likely voter" polls better than "registered voter" polls but, sometimes the "screen" is a little wacky.
Using simple expressed voter interest is a poor screen (better than nothing, but not a lot).
Using a demographic screen is slightly better than a voter interest screen, but obviously it is based on historical demographics which may not apply in a particular election.
Using a historical particpation screen is even better, but, some people will lie to you (and few polls have the money to put together a contact list of people who actually voted) and, it discounts new voters who were unable to legally vote in prior elections.
The best method is to blend the three methods, with greater weight being given to historical and lesser wirght to voter interest.
Oh, and BTW, I am amazed that Gallup would release this wacky result.