Mean-frequency scores of lexical sophistication are used to evaluate written and spoken language production. They are calculated using word frequencies extracted from a reference corpus. Using mixed-effects regression models, we analyse the strength of the relationship between L2 proficiency and mean-frequency scores in spoken and written texts using reference corpora representing different modes and registers. We control for task and topic effects. We observe that mean-frequency measures of lexical sophistication are considerably more influenced by the mode and register of the reference corpus used to calculate these scores than by language users’ proficiency level. Advanced language users produce more frequent vocabulary, typical of the target register, in both spoken monologues and written essays. These results provide evidence in favour of a conceptual and terminological shift from lexical sophistication to register appropriateness (as suggested by Durrant & Brenchley, 2019) to refer to the construct captured by mean-frequency scores of vocabulary use.