More than 84% of the links AI engines cite come from earned and third-party sources, which means the single largest input to ai citation sources is coverage you do not own and cannot purchase. That number comes out of 207 million citations, and it has held inside a narrow band for over a year.
Muck Rack published the fourth edition of its Generative Pulse study on October 28, 2026, drawn from more than 40 million AI responses and 207 million citations collected through 2026, plus a fresh September pull of 1,020 unbranded consumer prompts run across ChatGPT, Claude, Gemini and Perplexity. It is the largest public dataset on this question, and the most useful thing in it is not the headline share. It is the stability of that share across four editions, and the two findings underneath it that most teams are not acting on.
Where AI Citation Sources Actually Come From
Earned and third-party coverage supplies more than 84% of ai citation sources, with owned media at 15.2% and paid at 0.3%, so the mix an engine draws from looks almost nothing like the mix a marketing budget is usually split across. Across four editions the earned share has moved only between 82% and 89%.
The category breakdown is where it gets operationally useful. Third-party corporate content, meaning the review sites, directories, vendor comparisons and partner pages that talk about you rather than belong to you, rose from 24% to 25.1% and is now the single largest identifiable slice. Aggregators and encyclopedic sources rose from 17.4% to 18.3%. Government and NGO sites rose from 8.6% to 9.6%. Academic research rose from 4.0% to 4.4%. Every one of those categories grew, and journalism paid for all of it.
| SOURCE CATEGORY | PREVIOUS EDITION | OCTOBER 2026 EDITION | WHAT IT MEANS FOR A B2B BRAND |
|---|---|---|---|
| Third-party corporate | 24% | 25.1% | Review sites, comparisons, directories and partner pages are now the biggest slice. Being described accurately on other people's domains is the work |
| Aggregators and encyclopedic | 17.4% | 18.3% | Wikipedia-class references and roundups keep gaining. Entity consistency across them is cheap and compounding |
| Owned media | 13.7% | 15.2% | Your own site is growing but remains a minority input. Publishing more of it cannot substitute for the other 85% |
| Government and NGO | 8.6% | 9.6% | Mostly unreachable for a vendor, and a reason your absolute ceiling on some queries is lower than it looks |
| Academic research | 4.0% | 4.4% | Worth a named data study if you have real numbers. This is the slice original research buys into |
| Paid and advertorial | 0.3% | 0.3% | Flat for two editions. Sponsored placement does not convert into citation share |
Put the owned number next to the paid number and the strategy writes itself. Owned media is 15.2% and climbing slowly, which is worth having. Paid is 0.3% and has not moved. Everything else, roughly five sixths of the total, is somebody else's domain describing your category, and in most cases describing it without you.
Why Paid Placement Is Not An AI Citation Source
Paid and advertorial content accounts for 0.3% of ai citation sources and has not moved since May 2026, so sponsored placement is close to invisible to the engines even when it sits on a domain they cite constantly. The clearest proof in the study is a same-domain comparison rather than a cross-domain one.
That is the cleanest available answer to the question every enterprise marketing team asks in month one: can we pay to be in the answer. On this evidence, not reliably. The engines are not filtering out paid content by label so much as responding to the signals that editorial selection produces and sponsored placement does not: inbound links from unrelated parties, repeated citation by other writers, and the absence of promotional framing that makes a passage hard to lift as a fact.
This is also why the GEO conversation and the digital PR conversation have collapsed into one conversation. We laid out the mechanics of which signals decide who gets named in the anatomy of an AI citation, and the Generative Pulse numbers are the market-scale version of the same finding. Earned coverage is not a nice-to-have on top of a content program. For five sixths of the citation surface, it is the program.
The Journalism Decline Is A Supply Problem
Journalism fell from 27% to 22% of all citations, the largest single category shift across four editions of the study, and the decline shows up in every major engine rather than in one. ChatGPT went from 28.1% in January to 22% in September, with sharp drops in spring and July. Gemini declined gradually from 27.6% to 23.6%. Claude held up best, moving from roughly 28% to 24%.
Greg Galant, cofounder and CEO of Muck Rack, was direct about not celebrating it: we want to see more journalism in AI answers, not less. Read the sentence as a market signal rather than a lament. Five points of citation share moved out of news publishing and into corporate third-party content, aggregators, government sites and academic work. Those are the categories a vendor can influence, and they absorbed the entire decline.
The mechanism behind it is contested, and anyone selling you certainty here is guessing. Licensing is part of the picture: a separate study from OtterlyAI and Press Ranger found pages from OpenAI-licensed publishers averaged 10.2 citations per cited page against 6.9 for unlicensed pages, and the authors were careful to say the correlation is not proof that signing a deal produces visibility. Blocking is another part. Crawler policy, paywalls and the sheer volume of corporate comparison content being published all push the same direction.
Run The Same Prompt Twice And The Dashboard Breaks
When the study asked one prompt twice in a single day, ChatGPT repeated only 41% of the domains it had cited the first time and Gemini 45%, while Perplexity repeated 94%. That is the most expensive finding in the report, because it invalidates the way most teams currently measure AI visibility.
Share of cited domains that repeated when the same prompt was asked twice in one day (Generative Pulse, October 2026)
Think about what a 41% repeat rate does to a weekly report. If you check twenty prompts once a week and your mention count drops from nine to six, you have learned nothing. The expected variation from asking the same questions twice is larger than the change you just put in a slide. Teams have cancelled working programs on that noise, and teams have taken credit for it too.
The fix is not a better tool, it is a bigger sample and an honest confidence interval. Ask every tracked prompt at least five times per measurement window, report the mean share of voice with its spread rather than a single mention count, and never compare two engines on one pass each. We argued the same discipline for classic rank data in why your rank tracking numbers disagree, and the case is stronger here, because generative answers are sampled rather than ranked. Perplexity at 94% is the exception that proves the point: its retrieval is far more deterministic, so it needs fewer passes, and a single methodology applied flatly across all four engines will misreport three of them.
Crawler Access Is The Gate In Front Of All Of It
Among major publishers blocking Claude's search crawler, 81% received almost no Claude citations, which makes crawler access the precondition rather than one tactic among many. The same study found 12% of publishers that allowed the crawler also received almost no citations, so access is necessary and nowhere near sufficient.
Those two numbers together are the whole policy. Blocking is close to a guarantee of absence. Allowing is a ticket to the lottery. Most enterprise sites we audit are somewhere in between without having decided to be: a robots.txt edited by three teams over four years, a bot-management rule from a security review, and a CDN default nobody has read. The fastest check is also the dullest, and it belongs in a quarterly technical review rather than in a strategy deck.
Note the split between training agents and retrieval agents. A policy team that blocks training crawlers to protect intellectual property is making a defensible call. A policy team that blocks the retrieval agents has quietly opted the brand out of the answer surface, usually without knowing the two decisions were different. Separate them, write down which choice you made and why, and check the edge response rather than trusting the file, because a 403 from bot management overrides a permissive robots.txt every time.
Where To Put The Next Dollar Of GEO Budget
If earned and third-party sources carry more than 84% of citations and paid carries 0.3%, the allocation question answers itself: fund the work that gets other people to describe you accurately, and fund the access and measurement that lets you see it. Here is the order we run it in.
The uncomfortable part of this data for an agency to publish is that it argues against volume. Eight more blog posts move the 15.2% owned slice slightly. One accurate rewrite of how a widely cited review site describes your product moves the 25.1% slice, and the engines read that page far more often than they read your blog. We go further into the earned side of this in how to get cited by ChatGPT and the community angle in seeding Reddit threads that AI engines actually read.
Start with the crawler audit this week, because it is two hours of work and it gates everything else. Then pick the three third-party domains the engines cite most in your category and go and get described accurately on them. That is what the 84% is made of, and it is the part of generative engine optimization that does not happen on your own site.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.