Blog · 21 September 2026 · By HeardOf

Muck Rack's 84% is not 84% journalism. Here is what that number counts

What does Muck Rack's 84% actually count?

Everything that is not paid, owned or shared. The May 2026 edition of What Is AI Reading?, published by Muck Rack's Generative Pulse team on 7 May 2026, puts earned media at 84% of all citations, and the press release of the same date defines earned as "journalism, academic research, government sources, encyclopedic sites, and third-party corporate content". Journalism alone is 27%. The distance between those two figures is this post.

The number is not wrong. It is wide, and it is read narrowly. A communications team that hears "earned media is 84% of what AI cites" hears its own job description. In the same study Wikipedia is ChatGPT's top cited domain and PubMed Central is Claude's, and by Muck Rack's own definition both would count inside the 84%. No pitch gets you into PubMed.

What the shares look like, as published across three Muck Rack pages:

Two things about that table. The shares do not reconcile into one pie on any free page we read: if the 24% for corporate blogs and content sits inside the 84%, then journalism and corporate content together are 51% of citations and roughly a third of all citations, about two-fifths of the earned category, is academic, government and encyclopedic material; if part of that 24% is counted as owned instead, the split moves. Muck Rack does not publish the resolved breakdown outside the report. And the non-paid headline moved from "more than 95%" (July 2025) to "about 94%" (December 2025) to "about 99%" (May 2026).

Which engines were measured?

ChatGPT, Claude and Gemini, across 17 industries. Perplexity is not in the May 2026 edition, though Muck Rack's December 2025 edition did include it, and Google AI Overviews is not in the May 2026 edition either. Per engine, the May 2026 edition reports ChatGPT citing sources in 96% of responses at an average of five citations, Gemini in 82% at eight, and Claude in 55% at 13. The top cited domain is Wikipedia for ChatGPT, PubMed Central for Claude and Reddit for Gemini.

That matters for anyone importing the figure into a conversation about AI search generally. Our audit runs four engines, ChatGPT, Perplexity, Gemini and Google AI Overviews (from our own systems, as of 23 September 2026), and the run behind our own citation numbers, on 4 September 2026, ran two of them, ChatGPT and Perplexity. ChatGPT is the only engine both studies touch. Reddit is the clean example: Muck Rack reports Reddit as Gemini's top cited domain, while in our run of 4 September 2026 all 19 reddit.com citations came from Perplexity and none from ChatGPT. We did not run Gemini that day, so we have nothing to say about Muck Rack's Gemini result, and the May 2026 edition has nothing to say about Perplexity; Muck Rack's December 2025 edition did include it.

Why is owned media 13.7% there and 44.5% here?

Because the prompts are different, and the prompt mix is the likeliest reason for the gap. Muck Rack's own account of its method, published 12 May 2026, says it deliberately avoids transactional branded prompts — its example is "Can I watch HBO on Roku?" — on the grounds that such a query is a shortcut to the brand's own website, and builds instead from archetypes such as how-to, problem-and-solution and comparison, across thousands of companies in 17 industries; its results also report industry-trend queries.

Our run of 4 September 2026 asked buyer questions: 20 prompts in four B2B software categories, four of them head-to-head comparisons that name vendors. In that slice ChatGPT put 44.5% of its citations on a measured vendor's own site, 77 of 173, and Perplexity 9.2%, 36 of 391. Muck Rack's owned figure, over a far wider prompt mix, is 13.7%. The two measures are not defined identically: ours counts only citations to the six measured vendors in each category.

Muck Rack's own data predicts the gap. Its third takeaway, in Muck Rack's recap of its May 2026 webinar, published 27 May 2026, is that industry trend queries cite journalism in 46% of responses, more than twice the rate of how-to and comparative queries, and that how-to questions lean on reference and brand-owned material instead. So 84% is an average over a distribution of question types Muck Rack chose. A buyer deciding between four vendors is not asking the average question. Both figures can be right, and neither transfers.

Can a reader check it?

Not from the free pages. The Generative Pulse report page, read on 21 September 2026, is a landing page with three charts and a form; the report and the methodology it points to are behind that form, and we did not fill it in. The two blog posts that describe the method name no model versions for the May 2026 edition. The first edition's press release did name them — GPT-4o and 4o-mini, Gemini flash and pro, Claude sonnet and haiku, hundreds of thousands of queries run during July 2025 — but it does not say whether they ran through an API or the consumer apps.

Muck Rack reports ChatGPT citing sources in 96% of responses. What we can say is that nothing on the free pages tells you whether the 96% came from the API or from chatgpt.com.

There is also a plain inconsistency in the scale. Muck Rack's blog post of 7 May 2026 and its GlobeNewswire release of the same date both say more than 25 million links. Muck Rack's recap of its May 2026 webinar about the same edition, published 27 May 2026, says twice that the report analyzes 1 million or more links. That is a factor of 25, published by the same company in the same month about the same report, and the likeliest explanation is that the recap carried forward the December 2025 edition's "more than one million". We flag it because a reader quoting "1 million+" and a reader quoting "25 million+" are both quoting Muck Rack accurately.

Does it matter who published it?

Not by itself: vendor research is not disqualified by being vendor research, and the test is whether a reader can check it. Muck Rack sells Generative Pulse, a GEO monitoring product for communications teams, and the study's headline is that earned media drives 84% of AI citations. We are in the same position and will not pretend otherwise: we sell audits, and we publish audit numbers. The check is where the two differ. Our counting rules are published in who the AI actually cites, and our prompts, verbatim, in the post of 4 September 2026. Theirs are behind a form.

One more item, from Muck Rack's product announcement of 10 September 2026, read the same day as everything else here. Generative Pulse now uses millions of real consumer AI interactions to suggest which prompts a brand should monitor, and the post says that previously brands relied largely on prompts they created themselves or prompts generated from assumptions about what audiences might ask. That sentence is about the product, not the study. But the study's prompts were also constructed by Muck Rack, from its media database and a list of query archetypes, by its own description, and nothing we read says the May 2026 edition used observed consumer queries.

What should a team actually take from it?

Three findings survive a careful reading. Engines differ enough to be treated as separate channels, which is the most useful thing in the report. Query type shifts the source mix substantially: industry-trend queries cite journalism more than twice as often as how-to queries. And paid content is 0.3% of citations.

What does not survive is the inference most readers draw, that 84% earned means pitch more journalists. Journalism is 27%. The rest of the 84% is encyclopedias, academic and government sources, and third-party corporate content, and one lever Muck Rack's own data shows is not who you pitch but which question you are trying to be the answer to.

If the question you care about is a buyer's question, our run of 4 September 2026 says the source that answers it is your own page on ChatGPT and someone else's page on Perplexity. Neither of those is a press release.

How was this checked?

Every figure attributed to Muck Rack above was read on 21 September 2026 from pages anyone can open: the blog post of 7 May 2026 at muckrack.com/blog/what-is-ai-reading-may-2026, its GlobeNewswire release of the same date, the method post of 12 May 2026 at muckrack.com/blog/how-we-built-what-is-ai-reading, the webinar recap of 27 May 2026 at muckrack.com/blog/what-is-ai-reading-new-insights-may, the product announcement of 10 September 2026 at muckrack.com/blog/real-world-ai-search-data-to-geo-monitoring, the Generative Pulse report page at generativepulse.ai/report, and the first edition's release of 23 July 2025. The December 2025 edition's release of 2 December 2025 was read on 23 September 2026. We did not download the report itself, so every criticism above is a criticism of what Muck Rack published openly, not of the methodology inside the file, which may answer some of it.

Our own figures — 44.5%, 9.2%, 173, 391, 19 — are the four-category counts from our run of 4 September 2026 as re-derived from the stored answers on 19 September 2026, and they are set out in full, with the counting rules, in who the AI actually cites. That is the half of this post a reader cannot check independently, and we are saying so rather than letting you find the seam. Where we did not check something, the text says so.

Named but not cited: being mentioned by an AI is a different job from being its source