Species distribution models (SDMs) use spatial relationships between species occurrence and habitat (predictor) variables to generate maps of habitat suitability across a region of interest. These maps are frequently used in recovery planning efforts for endangered species, but they are influenced by data availability, selection of predictor variables, and choice of model type. Ground validation is necessary to robustly evaluate map accuracy, but it is rarely done, making it difficult to determine which modeling approach is best-suited for a given species or region. To address this uncertainty, we used two SDM types (Maxent and GLM) and two methods of selecting predictor variables to build four SDMs for an endangered prairie butterfly (Dakota skipper, Hesperia dacotae) in two regions of Manitoba, Canada. We then conducted field-based habitat suitability assessments at 120 locations in each region to enable direct comparisons of model output and accuracy. We found that soil type and surrounding landcover (grassland versus cropland) were important predictors of species occurrence regardless of region, predictor selection method, or model type. Cross-validation statistics indicated that most SDMs performed well (AUC > 0.7), but ground validation revealed that the habitat suitability maps they generated were inaccurate (Cohen's kappa < 0.4). Maxent models produced more accurate maps than GLMs, likely because false species absences adversely affected the latter, but only one Maxent-based map was accurate enough to help locate sites for future field investigations (Cohen's kappa > 0.3). Our results emphasize the importance of ground-validating SDM-based habitat suitability maps before incorporating them into species recovery plans.