They use the same subgroup, so long as they are participating, and circle back to them once a week. Each day has a sample size of somewhere between 300-400 people. Right now, the sample size is at about 2,200ish.
Apparently the folks at 538 claim they based their sample on the Obama/Romney race by asking who they voted for in that election, using the 51/47 breakdown for Obama/Romney in that race.
Of course the left-leaning hacks at 538 question that methodology.
If anything I would think that methodology would favor Clinton but who know.
Results are weighted to match demographic characteristics, such as race and gender, from the U.S. Census Current Population Survey and aligned to the 2012 presidential election outcome using self-reported vote in that election.
http://www.thegatewaypundit.com/2016/08/trump-top-latest-usc-la-times-daybreak-tracking-poll/