Journalistic content is integral to answers from generative AI tools, a new report from Generative Pulse by Muck Rack finds.
The authors of the report, published this week, analyzed more than one million citations output by generative AI models. Journalistic content was cited more than 27% of the time across all Muck Rack’s tests, and when it came to queries that implied “a level of recency” (including not just breaking news but timely responses to prompts like “car rental shortages in the U.S.” or “latest advancements in outpatient treatment methods for substance abuse”) that number jumped to 49%, or nearly half of all the links cited.
Muck Rack’s tests were conducted this month and included GPT’s 4o and 4o Mini models, Gemini’s Flash and Pro models, and Claude’s Sonnet and Haiku models.When ChatGPT and similar generative AI products craft responses, they can only rely on training data up to a point. Muck Rack’s tests confirmed that if you disable citations — and subsequently, the ability to scrape the web in real-time — for these models, they often output inaccurate or out-of-date information.
Muck Rack also found that the type of question asked changed the citation sources significantly. Subjective questions, such as queries asking for advice or step-by-step instructions, pulled far more from “corporate blogs and content.”
Journalistic content was cited in 49% of “queries implying a level of recency,” followed by corporate blogs and government/NGO sources.
Muck Rack also found some interesting differences between models. Overall, Claude appeared to cite journalistic sources the least often. Take Reuters as an example: The news source was cited by Claude 20 times less often than Gemini, and 50 times less often than ChatGPT.
The researchers also noted the top six media outlets cited by each model. Reuters, the Financial Times, Time, Forbes, and Axios all appeared in both the ChatGPT and Gemini lists. (It’s worth noting that the Financial Times, Time, and Axios all have AI licensing deals with OpenAI, but not with Google.) Again, Claude had less of a hard-news skew in its citations, with outlets like Good Housekeeping, TechRadar, and Harvard Business Review making its top-six list.
The Financial Times, Reuters, Time, Forbes, and Axios were among the most cited by ChatGPT and Gemini.
When it came to citations of journalistic content, the models seemed to favor stories that had been published in the last 12 months. That was especially true of ChatGPT, which saw 56% of its journalism citations published within the past year. For Claude, that number only came out to 36% of all citations.
The quality of citations in generative AI products have been a constant point of contention between AI developers and news publishers — particularly as generative AI tools take a slice out of traditional search traffic. ChatGPT and similar products have been found to hallucinate URLs to news publisher sites, even those with OpenAI licensing deals. They have also been found to link out to syndications and unauthorized copies of articles instead of the original stories. The frequency of journalistic citations, by the numbers, has been far less documented.
