This paper investigates the number of bootstrap replicates needed for accurate phylogenetic inference using Maximum Likelihood (ML). Phylogenetic bootstrapping (BS) is a method to estimate confidence in phylogenetic trees by generating many trees from slightly modified input data. The study focuses on ML-based BS and proposes stopping criteria to determine when enough replicates have been generated. The authors tested their criteria on 17 diverse real-world DNA datasets, including single-gene and multi-gene alignments with up to 2,554 sequences. They found that their stopping criteria typically stop computations after 100–500 replicates, producing support values that correlate better than 99.5% with reference values on the best ML trees. The results show that the number of replicates needed depends on the dataset, with some requiring several thousand replicates for the most conservative criteria. The study validates the proposed stopping criteria and shows that they significantly reduce computational costs while maintaining accuracy. The criteria are implemented in RAxML 7.1.0 and are available for download. The paper also discusses related work on bootstopping criteria and the convergence properties of BS values. The findings suggest that the number of replicates needed for accurate support values is highly dataset-dependent, and that adaptive stopping criteria can help balance accuracy and computational efficiency. The study contributes to the field by providing the first experimental assessment of the effect of BS replicates on support values and validating the proposed stopping criteria.This paper investigates the number of bootstrap replicates needed for accurate phylogenetic inference using Maximum Likelihood (ML). Phylogenetic bootstrapping (BS) is a method to estimate confidence in phylogenetic trees by generating many trees from slightly modified input data. The study focuses on ML-based BS and proposes stopping criteria to determine when enough replicates have been generated. The authors tested their criteria on 17 diverse real-world DNA datasets, including single-gene and multi-gene alignments with up to 2,554 sequences. They found that their stopping criteria typically stop computations after 100–500 replicates, producing support values that correlate better than 99.5% with reference values on the best ML trees. The results show that the number of replicates needed depends on the dataset, with some requiring several thousand replicates for the most conservative criteria. The study validates the proposed stopping criteria and shows that they significantly reduce computational costs while maintaining accuracy. The criteria are implemented in RAxML 7.1.0 and are available for download. The paper also discusses related work on bootstopping criteria and the convergence properties of BS values. The findings suggest that the number of replicates needed for accurate support values is highly dataset-dependent, and that adaptive stopping criteria can help balance accuracy and computational efficiency. The study contributes to the field by providing the first experimental assessment of the effect of BS replicates on support values and validating the proposed stopping criteria.