In the saddle point evaluation of the path integral, at tree level, you plug in the classical solution of the field into the integrand. However, when determining the classical solution, we ignore counterterms. The counterterms only show up to renormalize divergences after a saddle point expansion is made about the classical solution that does not include the counterterms. Why can we get away with this? It seems mathematically we have to include counterterms in the classical solution for a saddle point expansion to be valid.