My view on this question is similar to that of Eric Marcus (pdf).

When you think you're imagining a p-zombie, all that's happening is that you're imagining an ordinary person and neglecting to imagine their experiences, rather than (impossibly) imagining the absence of any experience. (You can tell yourself "this person has no experiences" and then it will be true in your model that HasNoExperiences(ThisPerson) but there's no necessary reason why a predicate called "HasNoExperiences" must track whether or not people have experiences.)

Here, I think, is how Chalmers might drive a wedge between the zombie example and the "water = H2O" example:

Imagine that we're prescientific people familiar with a water-like substance by its everyday properties. Suppose we're shown two theories of chemistry - the correct one under which water is H2O and another under which it's "XYZ" - but as yet have no way of empirically distinguishing them. Then when we epistemically conceive of water being XYZ, we have a coherent picture in our minds of 'that wet stuff we all know' turning out to be XYZ. It isn't water, but it's still wet.

To epistemically but not metaphysically conceive of p-zombies would be to imagine a scenario where some physically normal people lack 'that first-person experience thing we all know' and yet turn out to be conscious after all. But whereas there's a semantic gap between "wet stuff" and "real water" (such that only the latter is necessarily H2O), there doesn't seem to be any semantic gap between "that first-person thing" and "real consciousness". Consciousness just is that first-person thing.

Perhaps you can hear the sound of some hairs being split. I don't think we have much difference of opinion, it's just that the idea of "conceiving of something" is woolly and incapable of precision.