A ghost citation is an AI answer that links to your page but never names your brand. In a June 2026 study by Semrush, a vendor of AI-visibility tools, 61.7% of the 3,981 domain appearances it logged were ghost citations, and 13.2% were both cited and named. The sample is 115 prompts, which is small.
What the study counted
Semrush ran the study with Kevin Indig of the Growth Memo newsletter, using data from Semrush's own AI Visibility Toolkit. It logged 3,981 domain appearances for 115 prompts, run in 14 countries across four engines: ChatGPT, Google AI Overviews, Gemini and Google AI Mode (Semrush, 9 Jun 2026).
For each appearance it recorded two things: whether the domain appeared as a source link (cited) and whether the brand name appeared in the answer text (mentioned). An appearance is a domain showing up in an answer in either way.
Three outcomes
| Outcome | Share of appearances |
|---|---|
| Cited, brand never named (ghost citation) | 61.7% |
| Named, no citation | 25.1% |
| Cited and named | 13.2% |
Put differently, 74.9% of appearances carried a citation and 38.3% carried a brand mention. The citation rate is nearly twice the mention rate.
Insight
A citation tells the reader where a fact came from. The brand name is what a buyer remembers. Count them separately, because they move separately.
The engines do not agree
In this dataset Gemini named the brand in 83.7% of appearances and linked it in 21.4%. ChatGPT did the reverse: it linked 87% and named 20.7%. Google AI Overviews sat between them and leaned towards citations. AI Mode named brands at nearly twice ChatGPT's rate but still behaved more like a footnoted answer.
Across 454 prompt-and-domain combinations that several engines answered, the engines disagreed on whether to name the brand in 100 of them (22%). The lesson Semrush and Indig both draw is the same: visibility in one engine says little about another. A single blended score hides that.
Who gets ghosted
The most-cited domains were mostly aggregators and reference sites. medium.com was cited 16 times and never named; Wikipedia, Wired and harvard.edu also tended to be cited without being named. Brands with strong public identities went the other way: Google was named nearly three times as often as it was linked, Apple nearly twice.
So the 61.7% blends two different populations. A publisher should expect to be cited without being named. A consumer brand should expect the opposite, and should watch unlinked mentions on other sites as closely as links.
The shape of the question changes the answer
By query type, Semrush reports: informational queries (“what is”, “how does”) had an 89.3% citation rate and an 18% mention rate. Comparative queries (“best”, “vs”) had a 43.3% mention rate, 2.4 times the informational rate. How-to queries reached 42.8%. Commercial queries had a 35.6% mention rate and an 84.4% citation rate.
Short conversational prompts produced brand mentions almost every time; long structured prompts only 2% to 3% of the time, a 30 to 50 times gap on the same topic.
These are rates by query intent and prompt length. They are not an experiment on what you publish, so they do not show that a comparison page earns more mentions than a guide. Several summaries restate it that way. The study does not.
What one brand's run looks like
In the run behind our sample report (20 questions across four engines, 80 answers, fielded 8 Aug 2026), Company A was named in 44 answers. Its own site was cited in 34 of the 80.
That gives 27 answers with both a name and a citation to its own site, 7 with a citation and no name (ghost citations of its own site), and 17 named without a citation: 51 appearances, of which 7 (13.7%) were ghost citations.
This is one brand in one run. It shows how to read your own numbers. It does not show what is typical.
What this study cannot tell you
- It is a vendor's study on the vendor's own toolkit. Semrush sells the product it measured with.
- 115 prompts is a small panel. Rates by country, intent and prompt length rest on slices of it.
- The write-up does not say how a brand name was detected in answer text.
- Only domains that appeared at all are counted. It says nothing about the answers where you did not appear.
- It is observational. It does not show what makes an engine name a brand.
How to check your own
Run the same buying questions in short and long form on each engine, and for every answer record two separate things: was the brand named, and was its site linked. Do not average them. The method is in How to check what ChatGPT says about your business. The free Live Check runs one buying question on one engine.