About 84%, according to the most recent edition of Muck Rack's ongoing study of what ChatGPT, Claude and Gemini cite. The May 2026 edition of "What Is AI Reading?" analyzed more than 25 million links pulled from AI responses across 17 industries, and earned media accounted for 84% of them. Paid and advertorial content accounted for 0.3%. That headline number gets repeated constantly in PR decks right now, usually without the one detail that actually matters: Muck Rack built its own rules for what counts as earned, and some of those rules are not what most PR people would guess.
Where does the 84% figure actually come from?
Muck Rack has now run this study three times: July 2025, December 2025 and May 2026. Each edition pulls real AI responses, logs every source cited, and sorts them into three buckets, earned, owned and paid. The December 2025 edition, covering more than one million links, put earned media at 82% and non-paid sources overall at 94%. The May 2026 edition, with a much larger 25-million-link sample, moved the earned figure to 84% and held journalism specifically at 27% of all cited sources, a share Muck Rack says has stayed roughly flat across all three editions.
Muck Rack's CEO, Greg Galant, framed the practical implication directly in the May release: if your brand is not showing up in the media coverage AI is reading, you are not showing up in the answers AI is giving. That is a real finding. It is also one that depends entirely on how "the media coverage AI is reading" gets defined, which is where the classification rules come in.
What exactly does Muck Rack count as earned media?
This is the part most repetitions of the 84% skip. Muck Rack's research team answered the classification question directly in an FAQ its data lead, Matt Dzugan, gave alongside the study, and the rules are broader than a strict definition of earned coverage would suggest.
A paid contributor article, the kind published through a Forbes Council membership, counts as earned in Muck Rack's system. Asked directly whether that was right, Dzugan said yes: "in our categorization, that would show up as earned." That matters because Forbes Councils is a paid membership product, not press coverage a journalist chose to assign, and a strict reading of earned media would exclude it.
Affiliate content gets the same treatment, counted as journalism rather than paid, as long as the monetary relationship does not visibly shape the editorial content itself. Substack newsletters are also classified as journalism, even though Muck Rack's own data shows Substack accounts for less than 1% of citations, lower than the research team expected going in. Owned content, meaning a brand's own blog or newsroom, showed up so rarely in the sample that Muck Rack described its share only as "so low," without giving a number large enough to be worth stating precisely.
None of this means the 84% is wrong. It means "earned" in this study is closer to "not paid advertising and not your own website" than to the narrower definition PR practitioners usually mean when they talk about earned media: coverage a journalist or independent voice chose to create because it was newsworthy, not because anyone paid for the placement or the mention.
Does the platform you're trying to reach change what counts?
Yes, and this is the detail with the most practical weight. Muck Rack's data covers ChatGPT, Claude and Gemini, and the three models do not behave the same way. Gemini and ChatGPT draw from a similar pool of outlets. Claude is different: Dzugan notes it "has a tendency to avoid the bigger publications" that the other two models lean on. Outlets with hard paywalls get cited somewhat less often across all three models, which is a real cost for publications that gate their best reporting, and by extension for brands whose strongest coverage sits behind one of those paywalls.
There's also a quieter finding in Muck Rack's own recap of the research: across the brands it analyzed, only about 2% overlap exists between the journalists a brand's PR team pitches most and the journalists whose work actually gets cited by AI. That is a bigger operational problem than the earned-versus-paid split. A media list built around the outlets a comms team has always pitched is not necessarily built around the outlets these three models are actually reading from.
Why does the paid number stay so close to zero?
The 0.3% figure for paid and advertorial content is consistent with what Muck Rack's team describes as deliberate behavior on the model providers' side, not an accident of what happens to get crawled. Dzugan says AI systems are "being coached to avoid that sort of owned marketing and paid content," language that points to the model providers actively weighting training and retrieval away from promotional material. That's consistent with what we've seen elsewhere: ChatGPT and Google AI Mode now sell ads inside their answer surfaces, but the ad slot and the citation are two separate systems, and buying the first does not touch the second.
It also lines up with the AI Platform Citation Source Index, a separate 680-million-citation dataset covering five platforms, which found Reddit, Wikipedia and legacy news outlets dominating citations with essentially no role for paid placement. Two independently built datasets, measuring different things, arrive at the same structural conclusion: a model that cites sources is not going to cite an advertisement.
What should a PR team actually do with this number?
Three things follow from reading the classification rules rather than just the headline.
Stop treating the 84% as proof that any earned-adjacent content works equally well. A paid Forbes Council column and a feature a Wall Street Journal reporter wrote unprompted both land in Muck Rack's earned bucket, but they are not the same asset, and they are unlikely to carry equal weight with every model. What does a PR agency actually do covers the practical difference between the two and why conflating them backfires with clients who expect earned-grade results from a paid placement.
Check whether your coverage sits behind a hard paywall. If your best press mentions are locked, you are losing some citation value to outlets your buyers' AI tools can read freely, a tradeoff worth knowing about before deciding where to pitch next.
Audit the 2% overlap problem directly, rather than assuming your existing media list already covers the outlets AI models read. How to measure whether PR is actually working and what digital PR delivers that traditional PR doesn't both cover ways to check this without guessing.
The direct answer, for anyone who read this far
Earned media accounts for 84% of AI citations in Muck Rack's May 2026 study of over 25 million links across ChatGPT, Claude and Gemini, up slightly from 82% in December 2025, while paid content holds at roughly 0.3%. But Muck Rack's own definition of earned is wider than most PR teams assume: paid contributor columns like Forbes Councils count as earned, affiliate content counts as journalism, and only about 2% of brands' most-pitched journalists overlap with the ones AI models actually cite. The number is real. The classification behind it is the part worth reading before you repeat it.
If you want help figuring out which of your existing coverage is actually earning AI citations, request a quotation and we'll walk through what the data says about your category specifically.
