In approximate identification, the goal of the model should be taken into account when evaluating model quality. The purpose of this paper is the development of a system identification procedure, resulting in model sets that are suitable for subsequent robust control design. Incorporation of control relevance in the procedure results in a closed-loop frequency response-based multivariable system identification procedure. The model is represented as a coprime factorization, enabling the usage of stable model perturbations. The main result is the direct estimation of control-relevant coprime factors, exploiting knowledge of a stabilizing controller during the identification experiment. A numerically reliable iterative algorithm is devised, which is illustrated by means of experimental results.