This article claims to provide evidence on whether content-subject and cognitive achievement vary in CLIL (Content and Language Integrated Learning) and non-CLIL classrooms in the framework of a newly introduced CLIL programme in Italy. A standardized and validated science test (TIMSS) was administered to a sample 988 fourth-grade students. The sample was divided into three groups: one group learning science through English, one group through German, and a control group in which the students' school language, Italian, was employed. In our study the lack of taking into account causal relations which prevails in many prior studies is reduced by employing a counterfactual design. Results show a slightly negative effect of CLIL on science learning in both treatment groups in comparison to the control group. However, for both CLIL groups the negative effect exclusively pertains to the learning of purely content-related knowledge, while there are no differences between the non-CLIL and the two CLIL groups with regard to two other cognitive domains: application of knowledge and reasoning. Moreover, the paper shows significantly negative CLIL effects only among low language achievers and students from less advantaged parental backgrounds.