We focus on two real-space renormalization-group (RG) methods recently proposed for a hierarchical model of a spin glass: A sample-by-sample method, in which the RG transformation is performed separately on each disorder sample, and an ensemble RG (ERG) method [M. C. Angelini, G. Parisi, and F. Ricci-Tersenghi. Ensemble renormalization group for disordered systems. Phys. Rev. B, 87(13):134201, 2013] in which the transformation is based on an average over samples. Above the upper critical dimension, the sample-by-sample method yields the correct mean-field value for the critical exponent ν related to the divergence of the correlation length, while it does not predict the correct qualitative behavior of ν below the upper critical dimension. On the other hand, the ERG procedure has been claimed to predict the correct behavior of ν both above and below the upper critical dimension. Here, we straighten out the reasons for the discrepancy between the two methods above, by demonstrating that the ERG method predicts a marginally stable critical fixed point, thus implying a prediction for the critical exponent ν given by 2 1/ν 1. This prediction disagrees, on a qualitative and quantitative level, both with the mean-field value of ν above the critical dimension, and with numerical estimates of ν below the upper critical dimension. Therefore, our results show that finding a real-space RG method for spin glasses which yields the correct prediction for universal quantities below the upper critical dimension is still an open problem, for which our analysis may provide some general guidance for future studies.