A "ghost citation" is when an AI engine pulls a fact from your page and links it as a source, but never names your brand in the answer. The reader sees the claim and maybe a link icon, never who said it. A study Semrush published in June 2026, with Kevin Indig of Growth Memo, put a number on it: across four engines, roughly 61.7% of brand appearances were ghost citations: cited, not named. Citation and mention, it turns out, are two different outcomes that rarely travel together.
This is a reading of one multi-engine dataset, not a universal law. Every figure below is attributed, dated, and hedged. It sits alongside our two-lens visibility audit and the cited-isn't-recommended piece; together they map the ways a link can fail to become brand value. If you want the playbook rather than the diagnosis, skip to how to get named, not just cited.
What is a ghost citation, exactly?
A ghost citation is a source link with no brand mention attached. The engine used your page. It may even show a citation marker, but the answer text never says your name, so the reader learns the fact without learning who it came from. The Semrush study (with Kevin Indig / Growth Memo, published June 2026) logged 3,981 domain appearances across 115 prompts in 14 countries on ChatGPT, Google AI Overviews, Gemini, and Google AI Mode, using the Semrush AI Visibility Toolkit, and split every appearance three ways:
| How a brand appeared | Share of appearances | What the reader sees |
|---|---|---|
| Ghost citation. Cited, not named | ~61.7% | A fact and a link; no brand name |
| Cited and named | ~13.2% | The fact, the source, and who said it |
| Named, not cited | ~25.1% | The brand named in prose; no source link |
The single most common outcome. Nearly two in three appearances. Is the one that builds the least brand value: your page does the work, and someone else's answer takes the credit. All figures are vendor-reported and directional; treat the shape, not the third decimal, as the finding.
Do citations and mentions actually decouple?
Yes, and by more than most brands assume. Aggregate the table above and the two rates pull apart cleanly: the citation rate was about 74.9% (ghost + named-and-cited) while the mention rate was only about 38.3% (named-and-cited + named-only). A link is roughly twice as likely as a name. The citation and the total mentions are measuring different events, and the study shows they don't move in lockstep.
The starkest evidence is how differently the engines behave, which is the part worth internalising:
| Engine | Named in the answer | Cited as a source |
|---|---|---|
| Gemini | ~83.7% | ~21.4% |
| ChatGPT | ~20.7% | ~87% |
These are near-mirror images. Gemini tends to name brands in prose but rarely link them; ChatGPT tends to link sources heavily but rarely name the brand behind them. Measuring only one number on only one engine would give you a wildly wrong picture of your visibility. A pattern consistent with the low cross-engine overlap documented in do a few domains dominate AI citations. Per-engine, per-metric, or you're guessing.
Why does a model cite without naming?
Because retrieval and phrasing are separate steps. An engine can retrieve your page into its context, lift the fact, and then generate a sentence that states the fact plainly. Without the naming being a deliberate act of attribution. Three mechanics drive the gap:
- Paraphrase strips provenance. The model restates the claim in its own words; unless your brand name is part of the claim, there's nothing tying the sentence to you.
- The link is a footnote, not a sentence. Citation markers are attached after generation. The prose can carry a fact whose source sits only in a link the reader never opens.
- Naming is an entity decision. The model names an entity when it's confident the entity is the answer. A function of entity strength, not of whether your page was retrieved. Weak entity association means your facts get used anonymously.
The through-line: earning the citation is a content-and-reachability job; earning the mention is an entity job. They have different fixes, which is exactly why they decouple.
Does it vary by market?
Sharply. The same study found the mention rate swings by country: brands were named in about 50% of answers in India and Sweden, but only 18-22% in Italy, Brazil, and the Netherlands (Semrush/Indig, June 2026). The gap is large enough that a global brand can look well-recognised in one market and near-invisible in another off the same content. If you operate across markets, a single blended mention number hides more than it shows. Measure per market as well as per engine.
A citation is a link; a mention is your name. In a zero-click answer, the link is a footnote most readers never open, so being the uncredited source of a fact is closer to invisibility than to visibility. The work is making your name travel with your facts.
Why does this matter for AI visibility?
Because a ghost citation delivers almost none of the brand value a citation is assumed to. In the old model, a link earned a click and a visit; in a zero-click AI answer, most readers never click, so an unnamed link leaves no memory of you at all. You supplied the answer and stayed anonymous.
It also reframes what "we're cited a lot" means. A healthy citation rate with a weak mention rate is precisely the cited-but-not-named quadrant of our two-lens audit. A signal to stop optimising for more citations and start working the entity side. This doesn't contradict the case for earning citations; it refines it. Citations remain the supply that feeds answers. The point is that supply alone, without your name attached, doesn't build the brand, so both metrics belong on the dashboard, tracked apart. How to close the gap is its own playbook: how to get named, not just cited.
The durable read is simple. Measure your citation rate and your mention rate as two separate lines, on each engine and each market, and watch the gap between them. That per-engine, per-market split, and how it moves over time. Is what Buffy Intel is built to track. Questions: [email protected].