Abstract. Land evapotranspiration (ET) plays a crucial role in Earth's water–carbon cycle, and accurately estimating global land ET is vital for advancing our understanding of land–atmosphere interactions. Despite the development of numerous ET products in recent decades, widely used products still possess inherent uncertainties arising from using different forcing inputs and imperfect model parameterizations. Furthermore, the lack of sufficient global in situ observations makes direct evaluation of ET products impractical, impeding their utilization and assimilation. Therefore, establishing a reliable global benchmark dataset and exploring evaluation methodologies for ET products is paramount. This study aims to address these challenges by (1) proposing a collocation-based method that considers non-zero error cross-correlation for merging multi-source data and (2) employing this merging method to generate a long-term daily global ET product at resolutions of 0.1° (2000–2020) and 0.25° (1980–2022), incorporating inputs from ERA5L, FluxCom, PMLv2, GLDAS, and GLEAM. The resulting product is the Collocation-Analyzed Multi-source Ensembled Land Evapotranspiration Data (CAMELE). CAMELE exhibits promising performance across various vegetation coverage types, as validated against in situ observations. The evaluation process yielded Pearson correlation coefficients (R) of 0.63 and 0.65, root-mean-square errors (RMSEs) of 0.81 and 0.73 mm d−1, unbiased root-mean-square errors (ubRMSEs) of 1.20 and 1.04 mm d−1, mean absolute errors (MAEs) of 0.81 and 0.73 mm d−1, and Kling–Gupta efficiencies (KGEs) of 0.60 and 0.65 on average at resolutions of 0.1 and 0.25°, respectively. In addition, comparisons indicate that CAMELE can effectively characterize the multiyear linear trend, mean average, and extreme values of ET. However, it exhibits a tendency to overestimate seasonality. In summary, we propose a reliable set of ET data that can aid in understanding the variations in the water cycle and has the potential to serve as a benchmark for various applications. The dataset is publicly available at https://doi.org/10.5281/zenodo.8047038 (Li et al., 2023b).