Something of a bunfight has broken out in the comments of a post at Darren Naish's Tetrapod Zoology about the merits (or lack thereof) of differing methods of classification. My opinions on the main argument have been expressed before, and I'm still happy with what I said there. However, one of Darren's regular commentors, Jerzy, asked a question that I believe is worth discussing.
Why do we use evolutionary relationships as the basis of classification in the first place?
To those of us working in systematics on a daily basis, instilled with the concept that "nothing makes sense except in the light of evolution", this seems like a natural assumption. To a layperson (or even a non-taxonomic scientist), it may appear unnecessary and pointless. Evolutionary relationships are not the only basis on which we could build a classification - we could use overall similarity, for instance. Why is it wrong to classify falcons and hawks, or herons and storks, together if they are not related evolutionarily? After all, they still look a damn lot like each other, and they are still very similar ecologically. Doesn't it make sense to continue combining them?
The unfortunate truth is that any system of classification is going to be deficient in some way, simply because we are taking the whole gargantuan multi-dimensional field of biodiversity and attempting to distill it down to a two-dimensional hierarchy of taxa*. Combining falcons and hawks despite their convergent origins conceals their evolutionary relationships. Separating them obscures their ecological similarities. Which is the set of features that we most wish to express?
*Assuming, of course, that this is what you are going for. Another alternative would be to abandon purely hierarchical relationships and accept a system in which different taxa may overlap but neither one is a subset of the other, such as "photosynthetic organisms" and "unicellular eukaryotes". While we do this all the time in informal category labels, such systems have rarely been proposed as formal systems, probably because (A) there is then theoretically a nearly infinite number of categories that can be used simultaneously, and (B) people just don't like to think like that.
The important question here is what the purpose of the classification is. On the one hand, a classification is kind of a summary of everything we know to date about an assortment of taxa. If this was our sole purpose, we might argue that a classification by overall similarity (a phenetic classification) was the better option. However, a good classification is more than just a retrospective summary, it is also our source for future predictions. We don't just want the known features of a taxon, we also want some indication of what that taxon's unknown features are likely to be.
To use a basic example, the extinct Little Barrier Island snipe (Coenocorypha barrierensis, Scolopacidae) is known from only a single specimen caught on that island in New Zealand in the late 1800s (Miskelly, 1988). Apart from the features directly observable on that specimen, nothing is directly known about C. barrierensis, but we can infer a number of things from comparison with the taxa with which it is classified. For instance, we know that it probably laid eggs, because all other birds whose reproductive habits are known laid eggs. It was probably a ground nester, because other members of Scolopacidae (including other Coenocorypha) are ground nesters. It is arguably possible that Coenocorypha barrierensis was an arboreal species that reproduced like a coral by growing offspring as buds, because we can't directly prove otherwise, but we can say by inference that such a situation is highly unlikely.
It is in this predictive capacity that the evolutionary basis for classification comes into play. Phylogenetic relationships offer the sturdiest basis we have for inferring likely characters of incompletely known taxa - and needless to say, our knowledge of any taxon (including our own species) is always incomplete. In fact, the amount we don't know about a given taxon at any one time is always far in excess of what we do know. Classifications by overall similarity, on the other hand, only speak about what is known to date - they don't make any predictions about the future. True, phylogeny is far from perfect, and it is not invariably reliable - convergence and divergence do still happen, after all - but it is the most trustworthy in the long run. This is also why, if our understanding or opinions of phylogenetic relationships change, it is preferable for the classification to change with them, because that way we are always making our predictions on the best basis we can.