Artificial intelligence–enabled cheminformatics depends on transforming molecules into representations that models can compare and generalize from. Yet similarity induced by a representation is not equivalent to pharmacological similarity. This conceptual analysis examines how molecular encodings preserve, compress, or omit information; how biological context changes the meaning of similarity; and how these choices delimit defensible generalization claims. It develops a representation–context–claim account in which inferential strength depends on whether the representation retains information relevant to the property of interest, whether the biological context defining similarity matches the intended claim, and whether validation reproduces the transfer boundary implied by that claim. Activity cliffs, phenotypic profiling, heterogeneous bioactivity records, and out-of-distribution evaluation illustrate different ways in which representational proximity can overstate pharmacological continuity. The resulting position is neither that chemical similarity is uninformative nor that increasingly expressive representations are intrinsically superior. Similarity becomes epistemically useful when its representational basis, biological reference frame, and intended inference are aligned. Generalization in cheminformatics should therefore be treated as a calibrated claim about preserved information under specified contextual and distributional conditions.