Recurrent computations are ubiquitous in biological neural networks. The functional role of these recurrent computations has been studied recently with artificial neural networks. Artificial recurrent neural networks have shown performance advantages compared to feedforward architectures in visual object categorization, especially in challenging conditions such as in the presence of occlusion and clutter. We investigate the informational content of recurrent activations, hypothesizing that they might convey information about category-orthogonal auxiliary variables such as the location and orientation of the target object, to iteratively select information at subsequent timesteps to optimally discriminate object categories. We find that linearly decodable information about auxiliary variables increases across time in all layers, that this information is present in recurrent activations, and that altering it decreases task performance. These observations confirm the hypothesis that category-orthogonal auxiliary variable information is conveyed through recurrent connections and used for selecting information relevant for optimised category judgements.