From what I have seen, it is usually closer to 4% except in the largest polls. That said, Obama has consistently led nearly every national poll all year, so I disagree that hundreds of small leads do not mean much. A single polls margin of error may be 3-5%, even more with small local polls, but dozens of such polls showing the same winner make the overall outcome far more certain. We cannot quantify that greater certainty, and the different samples/methodology/weighting make the victory margin no more precise, but the simple question of who wins is fairly unambiguous.