Herein we propose the use of a consensus approach, CONSRANK, for ranking CAPRI models. CONSRANK relies on the conservation of inter-residue contacts in the analyzed decoys ensemble. Models are ranked according to their ability to match the most frequently observed contacts. We applied CONSRANK to 19 CAPRI protein-protein targets, covering a wide range of prediction difficulty and involved in a variety of biological functions. CONSRANK results are consistently good, both in terms of native-like (NL) solutions ranked in the top positions and of values of the Area Under the receiver operating characteristic Curve (AUC). For targets having a percentage of NL solutions above 3%, an excellent performance is found, with AUC values approaching 1. For the difficult target T46, having only 3.4% NL solutions, the number of NL solutions in the top 5 and 10 ranked positions is enriched by a factor 30, and the AUC value is as high as 0.997. AUC values below 0.8 are only found for targets featuring a percentage of NL solutions within 1.1%. Remarkably, a false consensus emerges only in one case, T42, which happens to be an artificial protein, whose assembly details remain uncertain, based on controversial experimental data. We also show that CONSRANK still performs very well on a limited number of models, provided that more than 1 NL solution is included in the ensemble, thus extending its applicability to cases where few dozens of models are available.