The importance of the inclusion of the 3d orbitals on third-row atoms in the correlation space in G2 theory has been systematically examined through calculations on the third-row G2 test set. Compared with standard G2, this G2(d) approach gives better agreement with experiment for the evaluation of ionization energies, a slightly poorer agreement for atomization energies, and much the same agreement for the very small sub-set of electron affinities and proton affinities. Overall, there is only slightly better agreement with experiment. However, when mixing of the 3d orbitals of the third-row atom with valence orbitals on the adjacent atoms is strong, inclusion of the 3d orbitals in the correlation space becomes a prerequisite to obtaining reliable results. Standard G2 theory is unsuitable in these circumstances. Similar conclusions pertain for the more economical G2(MP2)(d) method and for the full G2(QCI)(d) method. Inclusion of the 3d orbitals in the correlation space greatly increases the computer time required for a G2 calculation so some simple additive corrections to the G2 energy to approximate the effect of this inclusion have been investigated. These additivity methods generally underestimate the effect of the 3d orbitals but give reasonable agreement with the full G2(d) calculations in most cases. They cannot be used, however, in situations where the 3d orbital mixing is strong.