Online polls do not create a random selection among the populace and thus are of zero scientific value
I used to work as a survey supervisor and data analyst. All online polls can be ignored as mere noise. If a reputable outfit like Reuters puts their name to it, it should be taken seriously.
Now, when you look at recent polls at Real Clear ( http://www.realclearpolitics.com/epolls/latest_polls/president/ ) you'll see they jump around a lot. Like on June 30th one poll had Clinton up by 11 and another had Trump up by 4. Since the introduction of the cell phone, random statistical sampling has gotten real difficult. Reaching people by cell phone is unreliable, but calling people on ground lines means you over-sample older voters, which will skew results.
Most pollsters will thus put in more time trying to reach younger voters. But if you put in more effort reaching one demographic group inside your polling "universe", then you've destroyed the random-probability-selection that true statistical sampling depends upon to be accurate. It also costs more money and polling organizations are not terribly profitable.
So instead, most pollsters simply gather what samples they can and then statistically weight the results to reflect what they think the real picture is of the overall voting public. There's different demographic criteria to consider in deciding how you'll weight a batch of interviews within a given survey, so the result is a bunch of widely divergent results.
That is to say, in the past, like in the 80s and 90s when I was in the racket, polling results didn't differ so much. Most presidential polls stayed within the margin of error of each other, or never diverged more than a few points from that MoE. So clearly when there is a 15 point difference in spreads between two more or less parallel polls, you're in a new age of polling.
Random probability sampling has always been a messy science. Today it's more art than science.
= new reply since forum marked as read
