The neural network activated during Topographical Memory (TM) tasks in controls overlaps with the earliest affected regions in Alzheimer's disease (AD) but not with those of Semantic Dementia (SD). This suggests that clinical TM tests could be more bespoke to neural dysfunction in early AD and therefore more sensitive and specific. We hypothesized that TM impairment would be characteristic of AD but not of SD making it useful both for early diagnosis and differential diagnosis. TM was assessed in 69 patients (22 mild AD, 15 SD, 32 with mild cognitive impairment (MCI)) and 35 controls, using three tasks: the four mountains test and two novel tests in a virtual town (the Virtual Route Learning Test (VRLT) and the Heading Orientation Test). AD patients were impaired on all TM tasks. The VRLT was the most discriminatory; had the highest correlation with caregiver reports of navigation problems; and correlated strongly with memory, attention/executive function, and to a lesser degree, visuospatial ability. In contrast, SD patients performed well on the TM battery only becoming abnormal with very advanced dementia and performance correlated exclusively with attention/executive function. The VRLT achieved 95% sensitivity and 94% specificity in discriminating AD patients from controls; at the same cutoff, 70% of MCI patients were impaired. When combined with either naming performance or global dementia severity, there was complete separation of AD from SD. The VRLT is ecologically valid, highly sensitive to early AD, and useful in discriminating AD from the non-Alzheimer dementia, SD.