Within a prospective cohort study, we explored the long-term reproducibility of the food-frequency questionnaire (FFQ) and dietary changes in the European Prospective Investigation into Cancer and Nutrition (EPIC)-Heidelberg cohort. After a mean follow-up time of 68·8 (sd 4·1) months the dietary assessment by means of a validated FFQ was repeated in 21 462 participants in the EPIC-Heidelberg cohort. The correlation and test–retest reproducibility of both dietary intake measurements was explored. The long-term correlation coefficients ranged from 0·41 (vegetables in men) to 0·77 (alcoholic beverage consumption in women). The median intake of potatoes, added fat, sugar/confectionary, cakes and alcoholic beverages was lower in the second than in the first FFQ, whereas the median intake of fruits, vegetables, cereals/cereal products and non-alcoholic beverages were higher. Consistently for food groups, 60–70 % of the participants in both genders were re-classified to the same or adjacent quintile of intake. The results of fairly high correlation coefficients indicate good agreement between both measurements. It is acknowledged that this result reflects to a substantial extent the measurement error of the FFQ and conclusions on real changes in the diet should be drawn very carefully. For some nutrients the dietary changes were consistent with results from independent national surveys. The performance of the FFQ gives confidence in the dietary data to be used as long-term exposure variables.