What the AIs actually recommend, measured.
The AI Disagreement Index asks eight models the same buyer questions and scores how much they actually agree. The answer: moderately, and it falls apart on the frontier. A Fleiss κ of 0.41, mean pairwise agreement of 0.44, and models that contradict themselves about three times in four. Near-consensus on the mature categories, a five-way split on the emerging ones. It sits inside a 15-dataset portfolio data portal where every figure is open-licensed and carries a permanent DOI.