Taking the closest features of the last-found simplices (sets of support points) for the two objects tested using GJK gives you a rough approximation of the contact manifold but you can do better, since (a) this simplex is usually not the complete manifold and (b) it is not necessarily a face of the boundary.

The following single-shot method for computing the contact manifold does a better job:

Given a contact normal N and a support mapping A, the contact area can be defined as the set of support points

lim conv{S_A(U) : angle between N and U < alpha}
alpha -> 0

Thus, the contact area is the convex hull of the set of support points found by slightly perturbing the contact normal. You can perturb the contact normal by adding a tiny vector orthogonal to the contact normal. These tiny vectors can lie on a circle or can be generated randomly. For instance, you could generate the vectors by hierarchically slicing the unit circle until for some iterations no new support points are found. The actual contact manifold is the intersection (after projection) of the two convex areas.
This method is somewhat cumbersome, so in many cases Erwin's solution of maintaining the last-found contact in a persistent manifold is often sufficient and a lot cheaper.
Gino