<jats:p>Abstract. Reconstructions of past variations in the global mean surface temperature (GMST) are used to characterise the Earth system response to perturbations and to validate Earth system simulations. Beyond the instrumental period, reconstructions rely on local proxy temperature records and algorithms aggregating these records. Here, we propose to establish standards for evaluating the performance of such reconstruction algorithms. Our framework relies on pseudo-proxy experiments (PPEs). That is, we test the ability of an algorithm to reconstruct a simulated GMST, using artificially generated proxy data created from the same simulation. We apply the framework to an adapted version of the GMST reconstruction algorithm used in Snyder (2016) and the metadata of the synthesis of marine proxy records for the temperature of the last 130 kyr from Jonkers et al. (2020). We use an ensemble of four transient simulations of the Last Glacial Cycle (LGC) or the last 25 kyr for the pseudo-proxy experiments. Given the dataset and the algorithm, we find that the reconstruction is reliable for timescales longer than 4 kyr during the last 25 kyr. However, beyond 40 kyr BP, age uncertainty limits the reconstruction reliability to timescales longer than 15 kyr. For the long timescales, uncertainty on temperature anomalies is caused by a factor that re-scales near-global-mean sea surface temperatures to GMST, the proxy measurements, the specific set of record locations, and potential seasonal biases. Increasing the number of records significantly reduces all sources of uncertainty but the scaling. We also show that a trade-off exists between the inclusion of many records, which reduces the uncertainty on long timescales, and of only records with low age uncertainty, high accumulation rate, and high resolution, which improves the reconstruction of the short timescales. Finally, the method and the quantitative results presented here can serve as a basis for future evaluations of reconstructions. We also suggest future avenues to improve reconstruction algorithms and discuss the key limitations arising from the proxy data properties. </jats:p>