“Is That How Often We Got Named, Or How Often We Got Cited?”
Somebody asked me that across a meeting table while my AI visibility slide was still up on the screen. I’d built the chart myself. I couldn’t answer the question.
And I’d been showing that line as though its meaning were settled, without anyone once pushing on it, which is its own kind of warning. You’ve probably got a version of this chart too.
But a citation and a mention are two separate events, and drawing them as one line points next quarter’s work at whichever half happened to be getting measured. A study published in June 2026 sorted 3,981 of these appearances one at a time, across four AI platforms and 115 prompts, and the two halves turn out not to move together at all.
What’s The Difference Between Being Cited And Being Mentioned In An AI Answer?
Being cited means the AI listed your page or domain as a source. Being mentioned means your name shows up in the sentences the reader actually reads — and 62% of citations never get that far.
So the two events come apart in a very specific way: the link sits in the source list under the answer, while the answer itself talks about the category without ever naming the company. Kevin Indig, who worked on the study, calls that a ghost citation. It’s the co-author whose name appears only in the bibliography.
Which half matters more depends on what the answer is meant to do. Readers read the prose; the source list is a footer most of them scroll past. Count only the citations and you’ve measured the part of the page that gets the least attention.
- Citation: the AI put your page or domain in the source list under the answer.
- Mention: your name is inside the answer text, where the reader’s eye actually goes.
Both are worth having. But they’re not the same number, and a single line can only be one of them.
ChatGPT And Gemini Have Swapped Which Half They Hand You
ChatGPT is generous with sources and stingy with names: it produces citations in 87% of appearances, but the brand name lands in the answer text only 20.7% of the time. Gemini runs the other way round. It puts the name in the text in 83.7% of appearances and generates a citation in 21.4%. And Google AI Mode names brands at about twice ChatGPT’s rate.
So whichever assistant happens to get opened decides what turns up. Check ChatGPT, find nothing, conclude that the models don’t know you — and the engine that had been saying the name out loud all along goes unnoticed.
The study also lined up the same prompt against the same domain across engines, 454 pairs of them. In 100 of those pairs the engines disagreed: one named the brand, the other didn’t. That’s 22%.
More than one pair in five, on identical inputs. Since that much disagreement can’t be flattened into a single reading of the market, I’ve stopped treating any one assistant’s result as the industry picture.
Where You Can Still Move The Mention Number
The same study measured two things that sit on your side of the line rather than the model’s. Comparison content — head-to-heads, “X or Y” pieces, roundups with a verdict — produced 2.4 times the brand mentions that plain informational content did. And the shape of the question mattered enormously: mention rates differed by a factor of 30 to 50 between short queries and long ones.
A query is just the thing a reader types into the assistant. Ask it the short, conversational way — “Should I lease or buy a car for my business?” — and a brand name comes back almost every time. Load the same question up with context, constraints and instructions, and the mention rate falls to 2% or 3%. The long, structured prompt is where the citations arrive with no name attached to them.
So two moves come out of that, and neither of them is a rebuild:
- Write one genuine comparison piece, with a verdict in it.
- Write your headings in the short, conversational shape buyers actually type, and check whether your name comes back when the question is asked that way.
I started on the comparison piece, mostly because the other side of this is slow. Getting cited more usually means getting written about by somebody else — press, review sites, a third party with a reason to link — and that runs on other people’s calendars. For the citation side measured as a share rather than a raw tally, Bing’s citation share metric is the cleanest denominator anyone has published so far.
How Far Does This 62% Actually Travel?
Semrush ran this study, and Semrush sells AI visibility measurement. “Mentions and citations have to be measured separately” is a conclusion that happens to sell what they make, and I’d keep that in view while I read the number.
The sample is 115 prompts and 3,981 domain appearances across 14 countries. The same company publishes a separate index built on 126 million prompts; this is not that one, and it is by far the smaller of the two. And the article doesn’t say when the data was collected, which matters more here than it usually would, since AI search behaviour rewrites itself every few months and an undated snapshot could be describing an engine that no longer exists in that form.
Nobody’s reproduced it independently. Nobody’s run it in Japanese, or in any language outside those 14 countries. Maybe the ChatGPT-Gemini split isn’t about brands at all: it could be a difference in interface — whether the product chooses to surface citation links — rather than a difference in how the two models write. The study doesn’t separate display from generation, so I can’t rule that reading out, and if it’s the right one then the split says less about strategy than I’ve made it sound.
Although that leaves the 62% as a figure I wouldn’t paste into my own forecast, the structure underneath it is a different matter. “Citations and mentions move independently, and the direction flips by engine” is a claim you can test against your own data this afternoon, with no borrowed constants at all.
Give The Report Two Columns And One Row Per Engine
Tomorrow morning, open the visibility sheet you already keep.
Cut the count into two columns — times the name appeared in the answer, times the domain appeared in the sources — and give ChatGPT, Google AI Overviews, Gemini and Google AI Mode a row each. Numbers that had been added together will land in different cells, and some will refuse to land anywhere until you go back and check what you were counting. That refusal is the part I’d actually pay attention to.
Once they’re apart, the choice of next move makes itself. Sources but no name is a mention problem, and comparison content is the lever. If neither column fills up, that’s a citation problem, and the lever there sits with the press, the review sites and anyone else with a reason to link.
I’d rather have a sheet that admits it doesn’t know than a line that quietly averages two different things. One extra column in the sheet you already keep is enough to start.
Sources
- Margarita Loktionova (Semrush; contributions from Christine Skopec and Kevin Indig / Growth Memo), “Why 62% of AI citations don’t lead to brand mentions [Study]”, Semrush Blog, 9 June 2026 (limitation: Semrush sells AI visibility measurement tooling and therefore holds a commercial interest in the conclusion that mentions and citations must be measured separately. The sample is 115 prompts and 3,981 domain appearances across 14 countries. The article does not state when the data was collected. No independent third party has reproduced the result, and no equivalent measurement in Japanese has been published)
