<p>Retrieval practice (i.e., practice testing) enhances recall relative to other study methods, but it is still unclear when it helps most. Here, we ask if testing is more beneficial for weakly or strongly related word pairs. Participants studied word pairs with either low or high semantic relatedness, then practiced them using either test-with-feedback or restudy. We gave a final cued-recall test 24 hours later. Rather than relying on an ANOVA interaction on proportion correct, which can be hard to interpret when the link between latent memory strength and accuracy is unknown, we evaluated performance using a principled reference model and used cumulative distribution matching. Testing improved recall in both groups, but its benefit was 26% smaller for highly related pairs. This pattern was not explained by ceiling constraints, either across participants or at the item level. The model’s characteristic quadratic relation between test and restudy performance held, but the high-relatedness data showed a systematic reduction in the incremental gain from testing. One interpretation consistent with the model is that study and test memory strengths become positively correlated for highly related materials. These results clarify how semantic structure shapes the testing effect and when practice testing is likely to pay off most.</p>