The desert tortoise (Gopherus agassizii) was listed as a U.S. threatened species in 1990 based largely on population declines inferred from mark–recapture surveys of 2.59-km² (1-mi²) plots. Since then, several census methods have been proposed and tested, but all methods still pose logistical or statistical difficulties. We conducted computer simulations using actual tortoise location data from 21-mi² plot surveys in southern California, USA, to identify strengths and weaknesses of current sampling strategies. We considered tortoise population estimates based on these plots as “truth” and then tested various sampling methods based on sampling smaller plots or transect lines passing through the mile squares. Data were analyzed using Schnabel's mark–recapture estimate and program CAPTURE. Experimental subsampling with replacement of the 1-mi² data using 1-km² and 0.25-km² plot boundaries produced data sets of smaller plot sizes, which we compared to estimates from the 1-mi² plots. We also tested distance sampling by saturating a 1-mi² site with computer simulated transect lines, once again evaluating bias in density estimates. Subsampling estimates from 1-km² plots did not differ significantly from the estimates derived at 1-mi². The 0.25-km² subsamples significantly overestimated population sizes, chiefly because too few recaptures were made. Distance sampling simulations were biased 80% of the time and had high coefficient of variation to density ratios. Furthermore, a prospective power analysis suggested limited ability to detect population declines as high as 50%. We concluded that poor performance and bias of both sampling procedures was driven by insufficient sample size, suggesting that all efforts must be directed to increasing numbers found in order to produce reliable results. Our results suggest that present methods may not be capable of accurately estimating desert tortoise populations.