chatgpt citation study reveals alarming inaccuracies for publishers