<p>Cottonseed protein is a valuable yet underexploited resource. However, efforts to genetically improve this trait have been hindered by the scarcity of genetic loci that are reproducible across diverse environments. In this study, we evaluated the crude protein content in a natural population consisting of 259 upland cotton (<i>Gossypium hirsutum</i> L.) accessions across seven environments in Xinjiang during the 2023–2024 growing seasons and performed a genome-wide association study (GWAS) using 1,144,681 single-nucleotide polymorphism (SNP) markers derived from whole-genome resequencing. The results revealed a wide range of variation in protein content (27.00%–50.76%), a relatively high heritability estimate (H² = 0.68), and significant genotype-by-environment interaction effects, collectively underscoring the necessity of multi-environment association mapping. Using 12 phenotypic datasets comprising single-environment measurements and BLUP-derived integrated phenotypes, a mixed linear model identified 211 significant SNP–trait associations corresponding to 195 unique SNPs after removing recurrent detections across datasets. Among these, 11 SNPs were reproducibly detected in at least two datasets and were further consolidated into nine candidate association intervals distributed across six chromosomes. Haplotype analysis of the LD block surrounding D11:19,580,385, combined with functional annotation, enabled the prioritization of 13 putative candidate genes with annotation-level roles in lysine biosynthesis, protein translation, vesicle-mediated transport, ubiquitin-related regulation, and protein interaction processes. These genes warrant further functional validation in future studies. Collectively, the findings of this study provide recurrent association loci and a practical foundation for accelerating marker-assisted selection breeding and functional validation of genes underlying the cottonseed protein content.</p>