Now, some of this is plainly unfair. The LA Times, for example, polled the two early races of Iowa and New Hampshire, which produced significant errors from nearly everybody, and since that was two-thirds of their polling it makes them look pretty bad. Overall, in the six races they polled (NH, IA, CA), they were in the middle of the pack. SurveyUSA, by contrast, didn’t poll the early races at all, which almost certainly boosts their average.
For my money, it looks like Research 2000 might actually be the best of the bunch. They’ve polled 11 races, including the tough early ones, and have only been off by double digits once. Not bad.