Deletions are prevalent in the genomes of SARS-CoV-2 isolates from COVID-19 patients, but their roles in the severity, transmission, and persistence of disease are poorly understood. Millions of COVID-19 swab samples from patients have been sequenced and made available online, offering an unprecedented opportunity to study such deletions. Multiplex PCR-based amplicon sequencing (amplicon-seq) has been the most widely used method for sequencing clinical COVID-19 samples. However, existing bioinformatics methods applied to negative control samples sequenced by multiplex-PCR sequencing often yield large numbers of false-positive deletions. We found that these false positives commonly occur in short alignments, at low frequency and depth, and near primer-binding sites used for whole-genome amplification. To address this issue, we developed a filtering strategy, validated with positive control samples containing a known deletion. Our strategy accurately detected the known deletion and removed more than 99% of false positives. This method, applied to public COVID-19 swab data, revealed that deletions occurring independently of transcription regulatory sequences were about 20-fold less common than previously reported; however, they remain more frequent in symptomatic patients. Our optimized approach should enhance the reliability of SARS-CoV-2 deletion characterization from surveillance studies. Finally, our approach may guide the development of more reliable bioinformatics pipelines for genome sequence analyses of other viruses.