Similarity scores feel inaccurate and fail to reflect user effort, breaking trust in the system.