We have found that a search mutation is as good or better than regular random mutation in every case we have tested. It has been particularly noticeable that in some contexts it can be used to solve optimisations in a single iteration. This is because the quality of the building blocks is good. Rather than being random they give the best possible information at the time.
In other cases, where the function is not particularly amenable to parabolic fitting, the operator acts as creep mutation until approaching a minimum, where the parabola fitting ensures fast convergence onto the local optimum.