Consider the steps in a transition from one generation to the next, starting with a newly formed zygote:
When the transition from one stage to the next depends on genotype, then selection has occurred at that stage. Thus, to determine whether selection is occurring we construct expectations of genotype or allele frequencies at one stage based on the frequencies at the immediately preceding stage assuming that no selection has occurred. Then we compare observed frequencies to those expected without selection. If they match, we have no evidence for selection. If they don't match, we do have evidence for selection.
As we've already seen, it's conceptually easy (if often experimentall difficult) to detect and measure selection if we can assay genotypes non-destructively at appropriate stages in the life-cycle. What if we can't? Well, there's a very nice approach known as selection components analysis that generalizes the approach to estimating relative viabilities that we've already seen [1].