Two vendor studies published this year measured the same behavior: AI engines that cite a page as a source without naming the brand in the answer they generate. Writesonic, in a piece on Search Engine Land dated July 29, 2026, puts the rate at roughly 40% across seven AI engines. Semrush, publishing on its own blog on June 9, 2026, reported close to 62% across four engines, on a different sample size and a different engine mix.
What is a ghost citation?
A ghost citation is what happens when an AI engine links to a page as a source but never names the brand in the answer it generates. Writesonic defines it plainly: “An AI engine links to your page as a source but doesn’t mention your brand in the answer itself.” Semrush, in its own study of the same behavior, credits Kevin Indig with coining the term.
The Writesonic figure: around 40%, across seven engines
Writesonic analyzed approximately 16 million brand appearances over a 30-day period, across seven AI engines: Perplexity, Google AI Mode, Google AI Overviews, ChatGPT, Gemini, Grok, and Microsoft Copilot. The study’s headline figure, as stated in Writesonic’s ghost citation study on Search Engine Land: “Around 40% of AI citations didn’t name the source brand in the answer.” Search Engine Land credits the piece to Nikki Lam and Samanyou Garg. Garg is Writesonic’s founder, which is worth stating plainly, since this is vendor data about the vendor’s own market, published in a trade venue.
“a citation your reader never sees is visibility that only exists in a dashboard.”
— Nikki Lam and Samanyou Garg, Search Engine Land
Writesonic’s per-engine rates ranged from 19% on Microsoft Copilot to 52% on Perplexity, a spread that a single blended average does not show.
| AI engine | Ghost citation rate |
|---|---|
| Perplexity | 52% |
| Google AI Mode | 49% |
| Google AI Overviews | 41% |
| ChatGPT | 37% |
| Gemini | 25% |
| Grok | 22% |
| Microsoft Copilot | 19% |
The Semrush figure: 62%, four engines, 3,981 appearances
Semrush’s study, published June 9, 2026, ran a smaller and differently shaped test: 3,981 domain appearances across 115 prompts in 14 countries, on four engines only: ChatGPT, Google AI Overviews, Gemini, and Google AI Mode. Its own figures, as laid out in Semrush’s ghost citations study: “Almost 62% (61.7%) were ghost citations,” “Over 13% (13.2%) were both cited and mentioned,” and “Only about 25% (25.1%) were brand mentions without a citation.”
Semrush also reports: “74.9% of all brand appearances included a citation, but only 38.3% of appearances included a brand mention.” In our reading, that split, inside a single study, is the clearest evidence that citation and mention are not the same event counted twice.
Why the gap is the story, not either number
Writesonic’s approximately 16 million brand appearances and Semrush’s 3,981 domain appearances differ by roughly three orders of magnitude. The engine lists do not fully overlap either: Semrush’s four engines exclude Perplexity, Grok, and Microsoft Copilot, and two of those three mark the ends of Writesonic’s range, 52% on Perplexity and 19% on Microsoft Copilot. In our view, a behavior that one vendor’s count puts near 40% and another’s near 62%, on different samples and different engine mixes, is not a KPI yet. It is a diagnostic.
Two consequences for anyone tracking AI visibility
- Citations and mentions cannot share a column. A source link that never names the brand, and a brand named with no link, are opposite outcomes, and neither study measures what either one is worth. In our reading, Semrush’s own split makes the distinction concrete: a citation without a mention, and a mention without a citation, are different outcomes that a blended score would flatten into one number.
- An average across engines describes no single engine. Writesonic’s per-engine numbers run from 19% on Microsoft Copilot to 52% on Perplexity. Reporting the figure per engine keeps that spread visible. Averaging it away hides the exact detail an analyst needs.
What to do with two numbers that don’t match
Split any AI-visibility report into two tracked columns, cited and named, and break both out per engine instead of averaging them. Stop benchmarking your own measurement against either published figure: a number built on a different sample size and a different engine mix is not a bar your own tracking needs to clear. The gap between the two studies sits alongside a broader shift in AI-search sentiment covered here earlier this year. In both cases the discipline is the same: track your own inputs precisely rather than chase someone else’s headline number. If you are also tidying the structured-data layer on your own pages, our Schema.org JSON-LD generator is here.