I think the real problem with most uses of GAs (including this one) is that to the extent that they work at all, they're merely performing a random walk search driven by mutation, and the "crossover" operation is, for all intents and purposes, just functioning as a random-restart of the search, a jump to a new spot in parameter space rather than a useful shuffling of units of functionality.
Sometimes a random local search is exactly what the doctor ordered, of course, it's almost entirely trivial to code and especially if you tune the variance over time you can get pretty good results, though I'd agree that in general you can usually get better results in less computer time with a more typical algorithm (though you've still got to choose the algorithm, implement/integrate it, test it, etc., so you lose a lot of programming time unless it's already part of your framework).
Genetic algorithms should really shine when it's not clear how to cast the problem as a finite dimensional optimization problem, for instance if you're trying to evolve some constructive procedure to solve a problem rather than tweak variables to improve performance. You're absolutely right, aothman, most problems, even most problems specifically constructed to show off GAs, don't fit into this category, so there's no real point except that the algorithm sounds cool...