You can make an observation of a system whose state is undetermined. By interrogating the system for its state, a state becomes determined. Suppose, for example, that you have a flipped coin and sent it rotating in space, never hitting the floor. Is it heads or is it tails? Until it is looked at, the question doesn't really make sense. And for this case, we'll define "looking at it" to mean sticking out your hand and catching it. There is a probability that it's landed heads up in your hand, and a complimentary probability that it's landed tails up. Once it's in your hand though, you can confidently say which state it's in.
Now, you could definitely take issue with this example, because you could argue that the rotation of the coin is well described, so with initial conditions, you can predict its position at any given moment. But imagine a microscopic quantum system, and, for the sake of this simple explanation, believe that its "rotating through the air" state really does not have any precise heads or tails definition. Until something gets in the way of that system, creating an interaction that exchanges information about its observable state, it's not meaningful to say that it's in one of the observable states at all.
A superpositition of states, as such, is essentially the representation of a state in terms of a basis set of observables. In the case of the coin, heads and tails are the two observable states, they are orthogonal, and they fully represent the state space of the coin. You could flip the coin, and put its state vector into the form of sqrt(2)/2 * Heads + sqrt(2)/2 * Tails. This state isn't observable, but it can be described in terms of observable components, where the coefficients represent the probabilities that a given observable state will be measured upon observation.