Although mutation merits a separate box in the flow diagram, it is very much a background operator in most GA applications (as it is in nature). The purpose of the mutation stage is to provide insurance against the irrevocable loss of genetic information and hence to maintain diversity within the population. For instance, if every solution in the population has 0 as the value of a particular bit, then no amount of crossover will produce a solution with a 1 there instead.
In traditional GAs every bit of every solution is potentially susceptible to mutation. Each bit is subjected to a simulated weighted coin toss with a probability of mutation , which is usually very low (of the order of 0.01 or less). If mutation is approved, the bit changes value (in the case of a binary coding from 0 to 1 or 1 to 0).
There are schools of thought in the GA community which believe that mutation should only take place in solutions for which crossover was not approved or that only one mutation per solution should occur. Undoubtedly there are classes of problems for which each scheme is the most effective.