I get why that is. You need the mathematical definition to demonstrate why that is, and most "Big-O explainers" don't want to assume any significant amount of mathematical background. But, that definition isn't that hard. It's simply:
f(x) is O(g(x)) iff there exists a positive number M and an x_0 such that for all x > x_0, |f(x)| <= Mg(x).
And, if you're in an analysis of algorithms context, it's even easier, because you typically don't have to worry about this absolute value business.
Well, that M is essentially the reason you get to drop constant multiples of f(x). And, you drop the least significant terms of f(x) because g(x) dominates them, i.e. lim_{x -> \infty} g(x)/f(x) = 0. (No need to prove this, because this is what makes the "less significant" terms less significant.)
I would also like to add that the equals sign in f(x) = O(g(x)) is one of the most useful abuses of notation that I know of, but it can be misleading. It doesn't behave at all like a real equality because it's not symmetric, but it is transitive and reflexive. It actually acts more like set membership than equality.