We employed a triadic comparison task in patients with Alzheimer's disease (AD) and healthy controls to contrast (a) multidimensional scaling (MDS) and accuracy-based assessments of semantic memory, and (b) degraded-store versus degraded-access accounts of semantic impairment in Alzheimer's disease (AD). Similar to other studies using triadic comparison tasks, participants were asked to indicate which two out of three words (animal names) were most similar in meaning. Novel to this investigation, we contrasted performance on two semantic dimensions of strong and equal saliency to controls, but varying in their specificity (land/water versus bird/non-bird). Degraded-store accounts predict that the more specific bird/non-bird dimension should be more consistently impaired in AD, whereas degraded-access accounts predict that both dimensions, because they are equally salient, should be equivalently impaired in the disorder. The MDS results suggested that both patient and control group responses were not discriminable from random responding, consistent with previous studies. By contrast an accuracy-based analysis on the same data showed that controls showed good knowledge of both salient dimensions, and were evenly split in their individual preference for one dimension over another. In contrast, patients showed higher accuracy and sensitivity to the broader land/water dimension than to the more specific bird/non-bird dimension, consistent with a storage-based account of the semantic impairment in AD. Our results further suggest that MDS methods can fail to reveal important and systematic behaviour in semantic tasks, in both patient and control groups. (Contains 6 figures and 4 tables.)