You could say that treating all models of a first-order theory is a way of avoiding the arbitrary selection of a particular completion of that theory. There are other situations where it may be best to treat all completions - https://ncatlab.org/nlab/show/completion#nonunique.