There is a widespread belief that string loop massles divergences may be absorbed into a renormalization of σ-model couplings (space-time metric and dilaton). The crucial property for this idea to be consistently implemented to arbitrary order in string loops should be the renormalizability of the generating functional for string amplitudes. We make several non-trivial checks of the renormalizability by explicit calculations at genus 1, 2 and 3. The renormalizability becomes non-trivial at the log 2∊ order. We show that the log 2 ∊ counterterms are universal (e.g. the same counterterms provide finiteness both of two-loop scattering amplitudes and of the three-loop partition function) and are related to the log ∊ counterterms (β-functions) in the standard way dictated by the renormalization group. This checks the consistency of the Fischler-Susskind mechanism and implies that the renormalization group acts properly at the string loop level.