Semantic priming paradigms have been used to investigate semantic knowledge in patients with Alzheimer's disease (AD). While priming effects produced by prime-target pairs with associative relatedness reflect processes at both lexical and semantic levels, priming effects produced by words that are semantically related but not associated should reflect only semantic activation processes. This study was aimed at further investigating automatic semantic priming effects in AD patients when semantically related concepts with little to no lexical association are used. Twenty patients with mild to moderate AD and 20 matched controls (NCs) performed a lexical decision task on 30 concept pairs (15 in the living and 15 in the non-living domain) in an automatic semantic priming paradigm. In order to investigate the relationship between priming alteration and semantic damage, we chose concepts from a database. This allowed us to quantify semantic indexes relative to the structural representation at the feature level. No priming was found in NCs or mild AD patients, probably because feature similarity was insufficient in the concept pairs used. Similar to the hyperpriming observed in previous studies, the appearance of priming in the moderate AD group suggests early semantic damage in which attribute knowledge is partially affected. Furthermore, the finding that priming was predicted by the level of sharing (in the semantic system) of features common to the two concepts in the pairs indicates that the level of redundancy of attribute information is the main factor responsible for resiliency to neurological damage in AD.