“Which Study Did You Pull That Average From?”
I’ve put a single line from another company’s survey into a deck for executives. It was an average, the “ChatGPT cites N sources per answer” kind.
Then, in a meeting, that line couldn’t be squared with what someone else had brought from a different survey. Same name for the metric; one value about three times the other.
That gap is where I got stuck. Was one of the two surveys wrong? Or were both of them right?
The two tallies I’ve set side by side here: one counted more than 25 million links that appeared in AI answers, and the other counted 126 million prompts from the US.
How Many Sources Does ChatGPT Cite In One Answer?
It depends on whose count you read. By Muck Rack’s, 5; by Semrush’s, 15.
Muck Rack counted more than 25 million links contained in AI answers (the third edition of its report, published 7 May 2026). The assistants covered are ChatGPT, Claude and Gemini, three in all.
Semrush analyzed 126 million US AI search prompts (a prompt is one question sent to an AI) from January through April 2026 (Semrush’s announcement). Its four platforms are ChatGPT, Gemini, Google AI Mode and AI Overviews.
“Citations per answer” is how many sources an AI lines up in the course of one answer, and by name, both studies are reporting that same metric. Under that one name, one comes out at 5 and the other at 15, a threefold gap.
Put Gemini Next To ChatGPT And The Order Reverses
Here are both counts again, this time with Gemini in them:
- Muck Rack: ChatGPT 5 / Gemini 8
- Semrush: ChatGPT 15 / Gemini 3
On Muck Rack’s figures Gemini lists more sources; on Semrush’s, ChatGPT does. So which report you walked in holding decides which AI gets to be the one that “cites a lot.” A gap in the values alone might just be measurement noise. But once the order flips, the suspicion creeps in that the two aren’t measuring the same quantity to begin with.
How Far Apart Are The Counting And The Conditions?
As far as the published material lets me pick them out, the differences are these:
- Measurement period (Semrush states January to April 2026; Muck Rack doesn’t state one, and its report is the third edition, published May 2026)
- Platforms covered (three against four)
- Region (Semrush states the US; Muck Rack doesn’t state one)
- Unit of scale (25 million links against 126 million prompts)
Links and prompts don’t fit on one yardstick.
The link count tallied the sources that turned up in answers; the prompt count tallied how many questions got asked. It’s like lining up a diner that weighed the food it served with one that counted the orders it took, and filing both under “portion size.”
That’s the point where I stopped putting the two companies’ figures in the same table.
But neither publisher has disclosed how it decided what counts as one citation. So I can’t claim the four differences above are what caused the split; that causal link is my inference.
Where Does Another Company’s Average Belong In Your Deck?
Both figures were measured and published by the companies behind them. Muck Rack is a tool for PR and communications teams; Semrush is a company that sells this kind of AI visibility tracking.
I don’t think a stake in the outcome makes either figure a lie. But neither figure can claim more than “this holds under our own conditions.” And if that’s all either can claim, picking one and making it your target stops working, because the moment I pick one, I’m chasing an average whose period, platforms and region aren’t guaranteed to match ours.
So does the number have to stay out of your deck altogether? No, that part’s fine. Put “which study, from when, covering what” right beside the figure, and whoever reads the deck takes the conditions along with the number.
Decide The Three Parts Of Your Own Yardstick, Then Start Measuring
What counts as one citation? Which models get measured? How often does the measurement get run again?
I’d decide those three first and write them into a corner of the report under “measurement conditions,” because as long as last month’s numbers were taken under those same three rules too, the newer ones can sit beside them and the difference will mean something.
The other companies’ averages, I think, can come out of the target column. Since they weren’t taken under your conditions, whether you beat them or fall short, there’s no reading what happened.
The studies themselves don’t need throwing away. For a sense of where the industry sits, they work fine; all that changes is where they go, out of the target column and onto the map.
Sources
- Linda Zebian, “Earned media still drives 84% of AI citations. Here’s what that means for PR.”, Muck Rack Blog, 2026-05-07, https://muckrack.com/blog/what-is-ai-reading-may-2026
- Semrush, “Semrush Releases Expanded 2026 AI Visibility Index, Analyzing 126 Million AI Search Prompts”, 2026-06-26, https://www.semrush.com/news/463141-semrush-releases-expanded-2026-ai-visibility-index-analyzing-126-million-ai-search-prompts/
