CO2 retrievals with high quality facilitate resolving the sources and sinks of CO2 are helpful in predicting the trend in climate change and understanding the global carbon cycle. Based on a nonlinear least squares spectral fitting algorithm, we investigate the optimization method for CO2 products derived from ground-based high resolution Fourier transform infrared spectra. The CO2 vertical column densities (VCDs) are converted into column-averaged dry air mole fraction XCO2 by using the fitted O2 VCDs, and thus the system errors (e. g. pointing errors, ILS errors, zero-level offset) are corrected greatly. The virtual daily variation which is related to air mass factor is corrected with an empirical model. The spectra screening rule proposed in this paper can greatly improve the XCO2 quality. The CO2 retrievals before and after the optimized method are compared using a typical CO2 daily time series. After using the optimized method, the fitting error is reduced by 60%, and the two-hours-averaged precision is ~0.071% (equals to ~0.28 ppm), which is perfectly in line with the TCCON (the total carbon column observing network) threshold, i. e., less than 0.1%.