The major difference between the poly-A selection bias and the Ribo-zero bias is that the former can be evaluated and corrected. Both biases stem from the fact that the transcripts are partially fragmented in the first place, otherwise the entire mRNA would have been caught in the poly-A selection and the entire mRNA would have been excluded with the Ribo-zero kit. Using the pattern of slope along the gene in the poly-A library (like in Fig 4A) you can compute the fragmentation rate and get a corrected mRNA level which will be uniform along the mRNA using one parameter. In the Ribo-zero case, however, you don't know the probes they used so trying to correct the pattern is impossible.
It is true that the biases are reproducible but they will have an effect is several condition. For instance, if the amount of rRNA changes between two samples I think that the changes between different regions of mRNA will not be linear in the concentration of rRNAs, resulting with different bias for different genes, and when the fragmentation level is not uniform in two conditions, where the half-life of the mRNA is different, the effect will be more smear where then are less fragments and steep when there are more fragments.
