“Again, how do you call anything with 4 or 5% in?”
If you are serious, then think about it his way:
Imagine that there is some group voters answering you questions. If you were to talk to every voter you would get a count of answers (which might not exactly match an election for various reasons). If you were to randomly talk to half of those voters you would get a count that is very likely almost exactly the same as the answer you would get from talking to all voters. If you were to randomly talk to a quarter of the voters you would get an answer that is probably a little less accurate, but probably not to far from the complete group.
As you talk to a smaller and smaller percentage of the voters the chances of your result matching the complete interviews continues to fall.
The thing is that the chances or a random selection coming fairly close are much higher than most people would guess. mathematically those chances can be determined and the results of the small sample survey can be expressed by saying something like “there is a 95% chance that somewhere between 64.5 and 60.3 percent of the voters likely support Trump.” That means that the result can be expected to be wrong 1 out of 20 time, but 19 out of 20 times the complete set of voters can be expected to be between 64.5 and 60.3.
The logic you are presenting is sound but has some underlying assumptions.
One of them is that if people are lying to you they are doing so in a random manner.