“85.5% of AI citations come from earned media.” That line has been doing laps around LinkedIn for months, mostly posted by people who work in earned media, mostly with a tone of vindication.

It is not a plan. It is an average over a definition that the firms measuring it do not agree on, and the number in the same body of research that would actually change what your team does on Monday is a different one: roughly 2% overlap between the journalists your PR team pitches and the journalists AI models actually cite.

Those two figures come from different measurements. Merging them, which is what most of the LinkedIn posts do, produces a conclusion neither study supports.

Two numbers, two studies, one bad merge

The 85.5% traces back to a 5W Public Relations release dated May 15, 2026, which credits it to Muck Rack’s analysis of more than one million AI prompts. Go look at what Muck Rack publishes on its own blog and that exact figure is not there. July 2025 reported 89%. December 2025 reported 82%. May 2026 reported 84%, across more than 25 million links cited by ChatGPT, Claude and Gemini in 17 industries. The circulating number sits inside that band. It is a reading of a report, repeated until it hardened into a citation.

The 2% comes from somewhere else in the same research program. In the December 2025 edition, covering over a million links cited by ChatGPT, Claude, Gemini and Perplexity between July and December of that year, Muck Rack lined up two lists: the journalists PR teams pitch most, and the journalists AI cites most. Cofounder Greg Galant put it plainly: “only a 2% overlap, which shows the industry hasn’t fully adapted.”

Melissa Rosenthal, cofounder of Outlever, did the arithmetic nobody in the trade wants to say out loud: if the overlap is 2%, the industry is missing by ninety-eight points. That framing is hers, not Muck Rack’s, and it is the most honest way to read the finding. The channel works fine. The target list is the broken part.

Nobody agrees on what counts as earned

This is where the headline number falls apart on its own. Meltwater analyzed 5.35 million citations in April 2026 across eight models (ChatGPT, Grok, Claude, Perplexity, Gemini, Google AI Overviews, Google AI Mode and Copilot). It puts earned and news at 39.5%. Corporate sites, blogs, documentation and long-tail domains took 50.3%. Social came in at 7.2%. Press releases, 0.4%.

That is a forty-point spread against Muck Rack on the same phenomenon. Neither firm is lying. They made different calls about where a syndicated news item goes, where a technical documentation page goes, where a product review or a forum thread goes. Move one bucket boundary and the headline moves with it.

Notice that the two firms agree on the narrow slice where the definition is unambiguous. Muck Rack’s May 2026 edition puts journalism specifically at 27% of cited sources and paid or advertorial content at 0.3%. Meltwater’s press release number, 0.4%, lands in the same neighborhood. The fights are all in the wide middle: whether a corporate documentation page, a syndicated wire pickup or a Reddit thread belongs on the earned side of the ledger. That middle is most of the internet, which is why the headline number swings by forty points and the edges barely move.

I wrote about the adjacent version of this problem when I covered the statistical noise in AI visibility measurement: two vendors measuring the same brand can differ by more than 20 points on sampling method alone, before anyone argues about categories. With these citation studies you get both failure modes stacked, sampling and definition, which is why the aggregate percentage stops working as guidance.

Semrush makes the instability concrete. Across 230,000 prompts on three platforms over 13 weeks, from July 14 to October 12, 2025, how often ChatGPT cited Reddit fell from around 60% of responses in early August to around 10% by mid-September. On Google AI Mode and Perplexity, the same source held steady. When one domain can swing fifty points in six weeks on one model and not move on two others, an annual “where citations come from” percentage is a snapshot of a specific week, not a planning constant.

The one controlled test in the pile

Most of this research is observational. One piece is not, and it is the one people skip.

Stacker and Scrunch took eight articles and ran them through 944 prompt-platform combinations across five AI engines, under two conditions. Condition A: the article lives only on the brand’s own domain. Condition B: the identical article, same words, pushed out across hundreds of third-party news sites.

Condition A got cited 7.6% of the time. Condition B reached combined visibility of roughly 34%. The authors frame it as a 325% citation lift. In the ratio that matters for a budget conversation, the syndicated version was picked up about four times as often as the original sitting on the brand site, with no change to the text at all.

That result should bother anyone who spent a career in PR. The win did not come from an exclusive in a prestigious outlet. It came from the same paragraph existing in many places at once. The models are not rewarding the masthead. They are rewarding redundancy: the same claim, checkable in several independent places.

For a B2B company in Central America, this is more useful than the entire earned media argument. The trade press covering a narrow technical topic in Honduras, Costa Rica or Guatemala is thin. Chasing the kind of prestige placement that moves the needle in New York gives you maybe three shots a year, and two of them depend on knowing an editor personally. Distribution breadth does not work that way. It needs a piece worth reproducing and a route to places that will reproduce it, and both of those are things a four-person marketing team can build without a rolodex.

There is a corollary the content team will not enjoy. The classic announcement press release, the kind with nothing verifiable in it, stays close to invisible. Meltwater puts press releases at 0.4% of citations. Muck Rack puts paid and advertorial content at 0.3%. The format is not the problem. A release with no proprietary data point in it simply gives the model nothing worth quoting.

What IQ Source does with this

Our AEO work starts at the boring end of this argument, because the boring end determines everything downstream: writing down what counts as a citation before reporting a single percentage. If a visibility report says your brand appears in 30% of answers, that number is only usable when it is sitting next to how many times the query was run, on which platforms, on which days, and what was counted as owned versus third-party. Without that, you cannot tell a real gain from the ordinary variance of a model that never answers the same way twice.

The second piece is acting on the breadth finding rather than the headline. Concretely, that means prioritizing one piece with a real proprietary data point showing up on several independent domains over chasing a single prestige placement that costs far more of your team’s time. Same logic I used writing about how video content becomes a deposit AI can draw from: the unit that counts is not the view, it is the reproducible source.

The third piece is measuring your own version of the 2%. If your communications team keeps one contact list and the models are reading a different one, that distance is measurable today, brand by brand, and it is the most direct signal available for whether your press effort is building AI visibility or just producing clippings for the monthly deck.

None of this says PR stopped working. It says a channel can keep performing while the map you use to work it points somewhere else, and ninety-eight points of misalignment is too wide to keep assuming the old list is the right list.

Measure the gap between who you pitch and who AI cites