Campaign polls are least reliable the farther one is from an election, because most people (as in the current presidential race) are not really paying attention yet.
Polls also are unreliable in the transition from one reality in the news to a new one. The improving U.S. performance in Iraq is not yet reflected in polls, such as this one from Gallup, because of the lag time between changes in complex situations. When the changes are adequately reported and then, even more, when a new reality filters fully into the public mind, poll results begin to reflect the change. The development of a new poll consensus on Iraq will emerge, but only after time--assuming, of course, that prospects in Iraq do continue to get brighter for the Coalition and anti-terrorist Iraqis.
Part of the difficulty is that polls are very good at providing answers to specific questions, but not so good at measuring intensity of feeling or--to the point here--whether an opinion is fixed or fluid. Before an opinion given to a pollster shifts, it first softens, and that softening is often hard for survey takers to catch.
In economics, likewise, public opinioin measured in polls suggest that in a period where the economy first starts to deteriorate, people appear to be slow to realize it. And, they seem slow to shift their understanding again the next time the economy improves. You see it especially on employment figures and inflation, matters where an opinion poll majority can sometimes conflict with hard data.
People usually take their time in changing their assessments.