The very early steps of Au metal cluster formation in Er-doped silica have been investigated by high-energy resolution fluorescence-detected X-ray absorption spectroscopy (HERFD-XAS). A combined analysis of the near-edge and extended part of the experimental spectra shows that Au cluster nucleation starts from a few Au and O atoms covalently interconnected, likely in the presence of embryonic Au-Au correlation. The first Au clusters, characterized by a well defined Au-Au coordination distance, form upon 400 °C inert annealing. The estimated upper limit of the Gibbs free energy for the associated heterogeneous nucleation is 0.06 eV per atom, suggesting that the Au nucleation is assisted by matrix defects, most likely non-bridging oxygen atoms. The experimental results indicate that the formed subnanometer Au clusters can be applied as effective core-shell systems in which the Au atoms of the 'core' develop a metallic character, whereas the Au atoms in the 'shell' can retain a partially covalent bond with O atoms of the silica matrix. High structural disorder at the Au site is found upon neutral annealing at a moderate temperature (600 °C), likely driven by the configurational disorder of the defective silica matrix. A suitable choice of the Au concentration and annealing temperature allows tailoring of the Au cluster size in the sub-nanometer range. The interaction of the Au cluster surface with the surrounding silica matrix is likely responsible for the infrared luminescence previously reported on the same systems.