Random sampling continues at every stage as a new generation grows up. Here is an imaginary example:

Imagine a line of 100 pack-horses are walking single file along a hazardous mountain path: only 50 of them make it though; the other 50 fall off the path and crash down the ravine. It could be that the 50 survivors were on average genetically surer of foot than the rest; the sampling of 50 survivors out of the original 100 would then be non-random. Natural selection would be determining which horses survived and which died.

Alternatively, death could be accidental: it could happen whenever a large rock bounced down the mountainside from above, and the horses do not vary genetically in their ability to avoid the falling rocks. The loss of genotypes would then be random in the sense defined above. If we compared the genotype frequencies in the survivors and non-survivors, it is most likely that the two would not differ; the survivors would be a random genetic sample from the original population. They could, however, differ by chance. More AA than aa horses might have been unlucky with falling rocks and there would be some increase in the frequency of the a gene in the population.