Understanding the intrahost evolution of viral populations has implications in pathogenesis, diagnosis and treatment, and has recently made impressive advances from developments in high-throughput sequencing. However, the underlying analyses are very sensitive to sources of bias, error and artefact in the data, and it is important that these are addressed adequately if robust conclusions are to be drawn. The key factors include: (i) determining the number of viral strains present in the sample analysed; (ii) monitoring the extent to which the data represent these strains and assessing the quality of these data; (iii) dealing with the effects of cross-contamination; and (iv) ensuring that the results are reproducible. We investigated these factors by generating sequence datasets, including biological and technical replicates, directly from clinical samples obtained from a small cohort of patients who had been infected congenitally with the herpesvirus human cytomegalovirus, with the aim of developing a strategy for identifying high-confidence intrahost variants. We found that such variants were few in number and typically present in low proportions, and concluded that human cytomegalovirus exhibits a very low level of intrahost variability. In addition to clarifying the situation regarding human cytomegalovirus, our strategy has wider applicability to understanding the intrahost variability of other viruses.