Abstract Background The reproducibility of gene expression measured by RNA sequencing (RNA-Seq) is dependent on the sequencing depth. While unmapped or non-exonic reads do not contribute to gene expression quantification, duplicate reads contribute to the quantification but are not informative for reproducibility. We show that mapped, exonic, non-duplicate (MEND) reads are a useful measure of reproducibility of RNA-Seq datasets used for gene expression analysis. Findings In bulk RNA-Seq datasets from 2,179 tumors in 48 cohorts, the fraction of reads that contribute to the reproducibility of gene expression analysis varies greatly. Unmapped reads constitute 1–77% of all reads (median [IQR], 3% [3–6%]); duplicate reads constitute 3–100% of mapped reads (median [IQR], 27% [13–43%]); and non-exonic reads constitute 4–97% of mapped, non-duplicate reads (median [IQR], 25% [16–37%]). MEND reads constitute 0–79% of total reads (median [IQR], 50% [30–61%]). Conclusions Because not all reads in an RNA-Seq dataset are informative for reproducibility of gene expression measurements and the fraction of reads that are informative varies, we propose reporting a dataset's sequencing depth in MEND reads, which definitively inform the reproducibility of gene expression, rather than total, mapped, or exonic reads. We provide a Docker image containing (i) the existing required tools (RSeQC, sambamba, and samblaster) and (ii) a custom script to calculate MEND reads from RNA-Seq data files. We recommend that all RNA-Seq gene expression experiments, sensitivity studies, and depth recommendations use MEND units for sequencing depth.