What is a gauge transformation of the metric in GR? I'm confused by this question: 
Why can't General Relativity be written in terms of physical variables?
I can't quite see how you can make any change to the metric without either: (a) changing the physics; or (b) changing which point in your manifold represents which physical point in spacetime.  Either of these would, it seems to me, disqualify the change as being a gauge transformation, at least in the usual sense.
So is there some cunning way of changing the metric that avoids either (a) or (b), or is there some sense in which (b) can still be thought of as a gauge transformation?
 A: very smart people got this part really wrong. As you intelligently point out, EM is also invariant under diffeomorphisms, that alone does not make diff(M) a gauge group of EM. So we really need a better argument for saying that diff(M) is the gauge group of GR.
This has been pointed out before in the past, but i don't think there is a intelligent reply to these issues.
If we were really into finding a Lie group that would preserve GR physics, we might need to consider spacetimes which are allowed to have different curvature but test particles have the same apparent trajectories. I briefly speculate about such things in this question, but i don't think no one has take the time on these problems, me included.
A: Weyl, the inventor of gauge field theory, introduced the following notion, see his book pp. 121--123.  If you have a manifold $M$, which has an affine connection given by the Christoffel symbols $\Gamma$, but is not necessarily pseudo-Riemannian, then you can define a notion which is subtly different from a metric tensor, which is a conformal structure instead.  Then if you fix a gauge you can compare absolute values of tangent vectors at different points, but this notion of size might not be gauge invariant.  Naively writing down the $g_{\mu\nu}$ you get does not produce a tensor, it produces a « tensor of weight one » which twists by a scalar factor as you change the gauge.  I.e., it's not a tensor strictly speaking, but you can think of it as a twisted tensor (to me this is analogous to the sheaf of twisted differential operators which act on a line bundle but are not differential operators since they don't act on the functions on the manifold.) But if $M$ is pseudo-Riemannian (i.e., if the connection comes from a pseudo-Riemannian metric) you can choose a gauge for which $g_{\mu\nu}$ is a tensor.  
A: A diffeomorphism (the gauge transformations in GR) does move around the points in spacetime, but it also changes all of the fields, including the metric, leaving the physics unchanged. An example would be if you had some time function on spacetime $t$ and you did a "time evolution" diffeomorphism which mapped each surface of constant $t$ into the surface of constant $t$ that is 10 seconds to the future of the original surface. As long as you similarly map ("pullback/pushforward") the metric and all of the matter fields forward in time, there should be no way to physically distinguish the new configuration from the old one. Mathematically, though, you have changed the metric and the matter fields on the spacetime manifold, since the new fields on the $t=0$ slice are different than the old fields on the same slice, since the new fields at $t=0$ have the values that the old fields have at $t=-10s$.
A: One sense in which a change of coordinates is a gauge transformation is described here.
