We evaluated improvements made to the mandatory notification surveillance system for Legionnaires" disease in France by estimating its sensitivity in 1995 and 1998 using a repeat capture-recapture method. A case of Legionnaires" disease was defined as a person treated for pneumonia in whom legionella had been detected. Patient details were collected from (1) mandatory notifications; (2) the National Reference Centre for Legionella; (3) a postal survey of all hospital laboratories. The three sources were cross-matched and 715 individual cases were identified. A log-linear model, which included an interaction term between mandatory notifications and both the National Reference Centre and Laboratory sources, provided an estimated total of 1124 cases (95% CI 973-1275) in 1998, a twofold increase compared with 1995. The sensitivity of the surveillance system improved from 10% in 1995 to 33% (95% CI 29-38%) in 1998. Capture-recapture methods are important tools in the evaluation of surveillance systems.