Loading page…
Third-party sources still dominate, but news is losing ground to other categories. The study also finds large differences in which sources AI services cite—and how consistently.
Listen to this story
Muck Rack’s October study finds journalism made up 22% of AI citations in September, down from 27% in May, while earned and other third-party sources still accounted for more than 84%. The findings cover 40 million AI responses and 207 million citations from January through September 2026, but they reflect a vendor’s analysis and a changing snapshot—not a durable ranking. For publishers and communications teams, provider differences and inconsistent repeat answers make access and visibility worth tracking over time, rather than judging from a single response.
Government and NGO sources, encyclopedic sites, academic research, and third-party corporate content all gained share as journalism declined.
ChatGPT’s journalism share fell from 28.1% in January to 22% in September; Claude stayed roughly between 24% and 28%.
In same-day repeat tests, ChatGPT repeated 41% of cited domains, Gemini 45%, and Perplexity 94%.
Journalism supplied a smaller share of AI citations in Muck Rack’s latest study, falling from 27% in May to 22% in September. Yet sources outside brands’ own websites still dominated. The October 28 announcement puts earned and third-party sources above 84%, showing a shift within that group rather than a wholesale move toward company-controlled content.
The October edition of What Is AI Reading? is the fourth installment from Muck Rack’s Generative Pulse team. It draws on data covering more than 40 million AI responses and 207 million citations from January through September 2026. A separate September collection ran 1,020 unbranded consumer prompts across ChatGPT, Claude, Gemini and Perplexity.
The findings were published in a company press release distributed by GlobeNewswire. Muck Rack also sells Generative Pulse to communications teams seeking to understand and influence how AI services describe their brands. That commercial context matters: this is a vendor’s analysis of citation patterns, aimed partly at people trying to improve their visibility in AI answers.
Muck Rack’s earned and third-party category includes journalism, encyclopedic sites such as Wikipedia, government and NGO sources, academic research and third-party corporate content. Its share has ranged from 82% to 89% across the four editions. Journalism had held between 25% and 27% since July 2025 before the latest decline.
Most of journalism’s lost share went to other third-party categories, according to the study. The changes were spread across several kinds of sources rather than concentrated in one replacement for news:
Owned media—content brands publish themselves—also gained, rising from 13.7% to 15.2%. Product pages and practical guides were frequently cited for product and how-to questions. Paid and advertorial content remained a much smaller category: just 0.3% of citations, unchanged since May.
Journalism fell from 28.1% of ChatGPT’s citations in January to 22% in September, with sharp drops in spring and July. Several major publishers fell to near-zero citations in ChatGPT even though most still permitted OpenAI’s search crawler, the automated system that accesses pages for search. Other AI providers continued citing those publishers.
The provider comparisons were less uniform than the overall decline suggests. Claude’s journalism share stayed roughly within a 24%–28% band, while Gemini’s declined more gradually from 27.6% to 23.6%. Across the four services, only 212 domains appeared in every provider’s citations.
ChatGPT repeated 41% of cited domains when the same prompt was run twice in one day.
Gemini repeated 45% of cited domains in the same-day repeat test.
Perplexity repeated 94% of cited domains in the same-day repeat test.
The repeat test gives a practical warning to anyone checking whether a brand or publisher appears in AI answers: a single response can be an unstable snapshot. Muck Rack explicitly cautions against treating one answer as a measurement, given the differences it observed when asking the same question again.
Crawler permissions were another important distinction. Among major publishers blocking Claude’s search crawler, 81% received almost no Claude citations, compared with 12% of those allowing access. That is an observed association, not proof that permissions alone determine citation choices.
The study’s methodology also carries a time limit: AI citation behavior may change as systems are updated or retrained. Its findings describe the measured periods, not a fixed ranking of sources that publishers or communications teams can count on indefinitely.
Loading discussion...
Join the conversation
Explain when that preference would help—or get in the way.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.