Spatial unevenness in observation sites may hinder accurate global soil respiration (R-S) quantification. We quantified site representativeness using Voronoi volumes within an environmental feature space and implemented an iterative pruning strategy to identify a representative-optimal subset (1363 sites). Using process-based models as benchmarks, we developed a Representative-optimal Model (RM) to simulate global R-S dynamics (1982-2022). Compared to the Full-set model (FM), the RM significantly enhanced spatial consistency with benchmarks and improved alignment with FLUXNET observations and atmospheric CO2 fluctuations. Neglecting representativeness led to a 6.4 PgC year(-1) overestimation of global R-S (94.1 vs. 100.5 PgC year(-1)), while simultaneously underestimating its long-term trend and interannual variability. Arid zones contributed > 50% of this downward revision, reflecting corrected biases in water-limited regions. These findings demonstrate that optimizing existing dataset representativeness can effectively mitigate systematic biases and refine global carbon budget accounting. While expanding observational networks remains a long-term priority, we advocate routinely integrating spatial representativeness optimization into data-driven workflows to immediately correct structural sampling biases.