The discovery of the Josephson effect has for the first time given national metrology institutes (NMIs) the possibility of maintaining voltage references which are stable in time. In addition, the introduction in 1990 of a conventional value for the Josephson constant, KJ-90, has greatly improved world-wide consistency among representations of the volt. For 20 years, the Bureau International des Poids et Mesures (BIPM) has conducted an ongoing, direct, on-site key comparison of Josephson voltage standards among NMIs under the denominations BIPM.EM-K10.a (1 V) and BIPM.EM-K10.b (10 V) in the framework of the mutual recognition arrangement (CIPM MRA). The results of 41 comparisons illustrate the consistency among primary voltage standards and have demonstrated that a relative total uncertainty of a few parts in 1010 is achievable if a few precautions are taken with regard to the measurement set-up. Of particular importance are the grounding, efficient filters and high insulation resistance of the measurement leads, and clean microwave distribution along the propagation line to the Josephson array. This paper reviews the comparison scheme and technical issues that need to be taken into account to achieve a relative uncertainty at the level of a few parts in 1010 or even a few parts in 1011 in the best cases.