These sort of predictions take into account gerrymandering*. They also normally take into account voter participation rate, which includes voter suppression but politely weasels its way around mentioning it. Adjusting for voter participation rate is where most of the differing models, gut instincts, and spin enter the model - that part's legitimately hard and the people trying to deal with it are inherently biased. Do you use the historic rate even though there's pretty clear evidence for increasing voter suppression? How do you model that increase?

*Also lol @ the idea that the national election isn't gerrymandered. The state borders heavily favor conservative candidates, and neglecting to change that is the same as doing it intentionally in the first place.