A probabilistic linguistic multiattribute group decision-making (PLMAGDM) problem is studied from a reliability perspective based on an evidential reasoning approach and linguistic granulation optimization. As decision-making reliability, indicating the validity and accuracy, greatly depends on the reliabilities of both group similarity and the degree of familiarity, it is helpful to study a decision-making problem from the perspective of information reliability. To this end, a PLMAGDM method which considers reliability is put forward in this paper. First, two sets of information (i.e., probabilistic linguistic terms on alternative grades, and linguistic terms on familiarity) are given by experts as measures of expert reliability and attribute reliability. A granulation optimization model is then constructed to generate the utility of grades while maximizing group similarity, and a model that redistributes the probabilities of grades is established for the situation of incomplete information. We further provide another granulation optimization model to generate the degree of familiarity. Additionally, using the evidential reasoning rule, the aggregated weights of experts are obtained employing an optimization method that is used to retrieve attribute information. Lastly, through the method comparison in two cases as well as the sensitivity analysis, the validity and applicability of the method are verified.