Gaussian-type geminal basis sets for correlation energy calculations are usually obtained by optimization of the second-order energy functional. In the present work we propose a new strategy of generating the initial sets of geminals for these optimizations. Taking advantage of the information about the analytical structure of the second-order pair functions, this strategy produces reasonable starting values of the nonlinear geminal parameters, which lead to accurate second-order energies after only a short, crude reoptimization.