Earned media supplies 84% of AI citations, study finds
Muck Rack looked at more than 25 million links cited by ChatGPT, Claude and Gemini. Paid placements accounted for 0.3% of them. Your homepage is not where the machines do their reading.
A buyer asks an assistant to name three companies in your category. The answer takes seconds, and your website contributes almost nothing to it.
That is the finding behind a column by Boris Dzhingarov, published by Entrepreneur, which gathers the citation research on where models get their picture of a brand.
The largest number comes from Muck Rack, which analysed more than 25 million links cited by ChatGPT, Claude and Gemini. Earned media accounted for 84% of citations. Journalism alone accounted for 27%. Paid placements accounted for 0.3%.
Dzhingarov's reading of the spread is that trade publications and niche industry sites get pulled into category answers as often as famous mastheads. A company that will never make a national front page can still be cited.
The work that follows is old-fashioned. Reply to journalist requests quickly, give a writer a usable sentence rather than a brochure, and publish original numbers other people can repeat.
Second on his list is video and podcast transcripts. Ahrefs studied 75,000 brands and found YouTube mentions were the strongest single signal it measured, ahead of the traditional metrics.
The mechanism is unglamorous. Transcripts are text, and models read text at scale.
You do not need your own channel for this. Go on the niche shows your buyers already listen to, and say the company name out loud, clearly, more than once. A host who spends forty minutes saying "my guest's company" leaves nothing in the transcript.
Third are review platforms. Research by SE Ranking found that a small group of review sites dominates the review links cited in AI answers, even while their human traffic falls.
The practical version: claim your profiles on the two platforms your category actually uses, ask satisfied customers for reviews at a steady pace, and reply to what arrives. Dzhingarov's rule of thumb is G2 or Capterra for business software, Trustpilot and its peers for consumer brands.
Fourth are community threads. Reddit, Quora and industry forums now feed AI companies through licensing deals, and Dzhingarov rates this the slowest channel and the easiest to wreck. Post under your own name, answer what you can answer, and accept that seeded threads get removed by moderators and discounted by models.
Fifth is the reference layer: Wikipedia, industry wikis, directories and best-of lists. Those entries are written from press coverage, which sends the whole list back to the first item. One point costs nothing to fix — make every directory spell the company name the same way, because three variations read as three separate small strangers.
Your own site still has a job. It is where a machine verifies a claim, not where it finds one, so date the important claims and write them in sentences a machine can lift whole.
Dzhingarov suggests an audit that takes twenty minutes. Write down the five questions a buyer would ask an assistant about your category, run them through two different tools, and note where each cited source came from.
Then set that list against where last quarter's marketing budget went. On his account, for most companies the two barely overlap.