Satellite and surface observations of atmospheric carbon monoxide (CO) yield divergent emission estimates, yet the underlying causes of this discrepancy remain poorly understood. Here we present a multi-region analysis spanning 2015–2024 over China, North America, and Europe, assimilating MOPITT satellite retrievals and ground-based CO measurements from over 3300 monitoring stations. Our results reveal that the systematic underestimation of modeled surface CO is a common feature across all three continents, with median observation-to-model ratios of 1.82 (China), 1.65 (North America), and 2.25 (Europe). This bias progressively diminishes from urban to rural sites, by 44 % in North America and 58 % in Europe, and almost disappears at remote sites, demonstrating that representation error is an important driver of the surface-model mismatch. In contrast, comparisons of column CO between MOPITT and model show no systematic underestimation. Assimilation experiments further illuminate the structural nature of this divergence: while surface-data assimilation greatly improves agreement with in situ observations, it introduces a positive column bias relative to MOPITT. This bias is most pronounced over China and Europe but remains modest over North America, and accordingly the decadal trends from surface CO show a sharp decline in China (-5.92 % yr^-1) vs. plateaus after 2018 and 2016 in North America and Europe, while satellite-derived column trends are substantially weaker. Our multi-region analysis demonstrates that the satellite-surface discrepancy is a structurally embedded feature of the observing system, and advocates for joint utilization of both observation platforms in future inversion studies to achieve robust and self-consistent CO emission estimates.