“If We Want To Show Up In The AI, Can’t We Just Buy The Slot?”
I didn’t have an answer in the room. I said I’d look into it, which is what you say.
And it’s a fair question. We bought search ads to show up in search, so the same instinct arrives at the assistants and asks what the rate card looks like.
Except nobody has ever bought a slot in a ChatGPT answer. Whatever those links at the bottom of the response are, somebody else put them there.
So where do they come from? Somebody counted, in May 2026, across more than 25 million of them.
Where Do The Links In An AI Answer Actually Come From?
Almost all of them come from something a third party wrote.
Of more than 25 million links collected from ChatGPT, Claude and Gemini answers, 84% were earned media — news, academic work, government sources, encyclopedias and third-party company coverage (Muck Rack, published 7 May 2026). News outlets alone accounted for 27%.
Paid advertising and sponsored content came to 0.3%.
The counting was done by Muck Rack, which sells a media database and measurement tooling to PR and communications teams, and the tally runs across 17 industry verticals.
Earned media, if the term hasn’t come up in your building, means the coverage nobody paid for. Someone under no obligation to you read about your thing and decided it was worth writing up, which is exactly what makes it the one kind of coverage you cannot purchase directly.
Is 84% Just One Month Being Weird?
That was my first move on it, because a single reading of anything is a coin flip you’re allowed to reject.
The same study has run in three editions since July 2025. Earned media lands between 82% and 89% across them, and news on its own between 25% and 27%.
Three separate readings, one narrow band. Not a month with unusually loud news in it.
That distinction is the whole reason the number is usable internally. One edition gets pushed back on in the meeting and deserves to be; three editions holding the same band is the sort of thing you can actually move a budget line against, and the difference between those two situations is most of what “evidence” means in practice.
The Models Don’t Cite Alike, Not Even Close
Three assistants, three habits, and the spread is wide enough to change what a single test tells you.
- ChatGPT attaches citations to 96% of answers, averaging 5 per answer
- Gemini attaches them to 82%, averaging 8
- Claude attaches them to 55%, averaging 13
Put the same question to two of them and one comes back with a stack of sources while the other comes back with prose and nothing underneath it.
Which matters for what you conclude. Open one assistant, fail to find your name, and the thing you have measured is that model’s citation habits rather than your own visibility.
You tasted one course and reviewed the restaurant.
So the count runs across several models at once, and the per-model citation rates get established before anybody reads anything into the totals.
So Which Column Does The Budget Go In?
Read the figures plainly and AI visibility drifts over to the communications side of the house.
What gets pulled into answers is reporting, trade press and third-party comparison pieces. That is where the work has to land, which means it lands on the people who already talk to writers.
Concretely, it starts with putting your own first-hand material into a shape a journalist or an analyst can pick up and use: press releases with actual figures in them, research where the method is written down, case studies with named customers and real numbers.
A PDF sitting on your own site, that nobody else has ever written about, is not going to be cited. From where the model sits, it may as well not exist.
For which surfaces are getting pulled in beyond the press, there are numbers on that in the piece about the models preferring Reddit threads to your official site.
How Far Does This 84% Travel?
Not all the way to a target on a slide. Five things sit between the figure and your goal sheet.
- Muck Rack sells tooling to PR teams, so “earned media is where AI citations live” is the conclusion that sells the product
- The publisher doesn’t state its limits. The prompt set, the extraction rules and the measurement period for each edition aren’t disclosed
- The answers are English-language. Nothing here measures how citations are composed in other languages
- “Earned media” is defined broadly enough to include encyclopedias and government sources, so 84% is not the share your communications team could go and win
- This counts citations. It doesn’t measure how much any of those citations moved a buyer toward choosing you
The fourth one is the one that bites when targets get set. Nobody has published what fraction of the 84% is Wikipedia and government pages, so a comms team that writes 84% onto its goal sheet is chasing a number that was never available to it.
The surfaces being cited were not the paid ones. None of the five caveats touches that part.
That holds whether the true figure is 82% or 89%, and whether the definition is drawn tightly or loosely. Five caveats and you can still have the budget conversation.
Move AI Visibility From The Ad Column To The PR Column
There’s one thing to shift this week, and it’s a line in a budget.
Hunting for AI ad inventory can stop. That inventory is a surface holding 0.3% of the citations, and no amount of headcount pointed at it changes the share.
What that effort goes into instead is being written about by other people — giving reporters, trade press and comparison writers first-hand material they can use. Which is work your communications team already knows how to do. No new department, no new tooling.
Then count whether your name actually got pulled in, across several assistants rather than one, because of what the citation rates above do to a single-model reading.
Start by moving the line. Advertising column to communications column, one row, this week.
Sources
- Muck Rack, “What Is AI Reading?”, published 7 May 2026 (more than 25 million citation links collected from ChatGPT, Claude and Gemini answers across 17 industry verticals. Earned media — news, academic, government, encyclopedia and third-party company sources — accounted for 84% of links, news outlets alone for 27%, and paid advertising and sponsored content for 0.3%. Across the three editions published since July 2025, earned media ranges from 82% to 89% and news alone from 25% to 27%. Citation behaviour by model: ChatGPT cites in 96% of answers at an average of 5 links per answer, Gemini in 82% at an average of 8, Claude in 55% at an average of 13. Muck Rack sells a media database and measurement tooling to PR and communications teams. The prompt set, extraction rules and per-edition measurement periods are not disclosed, and the answers measured are English-language.)
