Recent developments in high-throughput sequencing (HTS), also called next-generation sequencing (NGS), technologies and bioinformatics have drastically changed research on viral pathogens and spurred growing interest in the field of virus diagnostics. However, the reliability of HTS-based virus detection protocols must be evaluated before adopting them for diagnostics. Many different bioinformatics algorithms aimed at detecting viruses in HTS data have been reported but little attention has been paid thus far to their sensitivity and reliability for diagnostic purposes. Therefore, we compared the ability of 21 plant virology laboratories, each employing a different bioinformatics pipeline, to detect 12 plant viruses through a double-blind large-scale performance test using 10 datasets of 21- to 24-nucleotide small RNA (sRNA) sequences from three different infected plants. The sensitivity of virus detection ranged between 35 and 100% among participants, with a marked negative effect when sequence depth decreased. The false-positive detection rate was very low and mainly related to the identification of host genome-integrated viral sequences or misinterpretation of the results. Reproducibility was high (91.6%). This work revealed the key influence of bioinformatics strategies for the sensitive detection of viruses in HTS sRNA datasets and, more specifically (i) the difficulty in detecting viral agents when they are novel or their sRNA abundance is low, (ii) the influence of key parameters at both assembly and annotation steps, (iii) the importance of completeness of reference sequence databases, and (iv) the significant level of scientific expertise needed when interpreting pipeline results. Overall, this work underlines key parameters and proposes recommendations for reliable sRNA-based detection of known and unknown viruses.