Ghost Citations: When AI Cites You Without Naming You

A citation looks like a win in every dashboard. Then you read the answer and your name is nowhere in it. Semrush and Kevin Indig examined 3,981 domain appearances across 115 prompts, 14 countries, and four AI engines, and found that 61.7% were citations with no brand mention: the page was linked as a source, and the answer never said who wrote it. Only 13.2% of appearances got both. Indig's name for the pattern, ghost citation, has stuck because it describes something most measurement setups cannot see — they record one visibility signal where there are two.

Key takeaways

Two independent things can happen when an AI engine answers a question your brand is relevant to. It can write your name into the answer. It can link one of your pages as a source. Neither implies the other, which gives four outcomes rather than two.

Most tools resolve this into a single flag: visible or not. That is the mistake, because the middle two rows have opposite fixes. Named-only says your content is not being retrieved, so structure, freshness, and crawlability are the levers. Ghost-cited says the retrieval works and the attribution does not, which no amount of heading tidying will change.

The third row of that table is the one with no obvious lever, because attribution is not something on the page. A model that lifts your finding and presents it as a fact about the world has treated the fact as belonging to nobody. Indig describes the same decoupling running the other way, in the quarter of appearances that are a name with no source: "The AI knows the information about the brand came from somewhere, but doesn't feel the need to explicitly say so to users. The brand name carries on its own." Either direction, the engine is deciding independently whether to hand over the link and whether to hand over the name.

Search for a ghost citation rate and you will find 61.7%, 73%, and a widely repeated "about 28%" all presented as the number. They are not measuring the same thing, and one of them is measuring the opposite thing.

Read the third column and the range stops being contradictory. AirOps counts the overlap rather than the gap, and counts it over answers rather than over domain appearances, so its 28% is not a rival estimate of Semrush's 61.7% — it is a different measurement. A figure with no stated method, meanwhile, cannot be checked at all. Semrush's is the one to quote because it publishes what it divided by; AirOps' report explicitly leans on "millions of datapoints" without disclosing prompts, brands, or period; Superlines' 73% is a vendor number about its own tracking.

What survives all three is the shape: the overlap between being cited and being named is small, and every independent attempt to measure it lands somewhere between a fifth and a third of appearances. That is enough to act on. It is not enough to put a specific percentage in a board deck.

The per-engine split in the Semrush data is the most actionable part of it, because the two leading engines fail in mirror images of each other.

On ChatGPT, you will mostly be a link. On Gemini, you will mostly be a name with no link. Google AI Mode named brands roughly 40% of the time, close to twice ChatGPT's rate. Anyone reporting a blended "AI visibility score" across these engines is averaging two opposite behaviours into a number that describes neither.

Prompt intent moves it as much as the engine does. Informational prompts in the same study cited at 89.3% and mentioned at 18%. Comparative prompts named brands 43.3% of the time, about 2.4 times the informational rate — which makes sense, since an answer comparing options has to name the options. Language mattered too: mention rates reached about 50% in India and Sweden against 18–22% in Italy, Brazil, and the Netherlands. If your tracked prompt set is mostly how-to questions, a low mention rate is partly a property of the set, not a verdict on the brand.

Seer Interactive analysed 541,213 LLM responses across 20 brands and six platforms in February 2026, using Scrunch AI data, and found the two signals strongly linked: when a brand was mentioned, its citation rate was 53.1%; when it was not mentioned, 10.6%. A five-fold difference.

Be careful with the direction of that arrow. Nobody has shown that earning a mention causes a citation. The likelier reading is that both are downstream of the same thing — how strongly the model associates your name with the category — so they rise and fall together. That is still useful, because it means the work that fixes one tends to fix the other, and it explains why purely on-page changes move neither.

AirOps adds the durability angle: brands earning both a mention and a citation were 40% more likely to resurface in consecutive runs than citation-only brands, in a landscape where only 30% of brands stayed visible across back-to-back answers and 20% survived five consecutive runs. That matches what we measured directly in our own 42-day citation study, where the set of domains cited for a prompt turned over 60–70% day to day and only a small stable core appeared every day. Ghost citations, in other words, are also the least durable kind of appearance.

Seer's sharpest illustration is a single client: a compliance software page that earned more than 100 citations in 25 days without a single brand mention. Perfect retrieval, zero recognition.

Three limits worth stating before anyone rebuilds a strategy on this.

The samples are small or undisclosed. 115 prompts across 14 countries is roughly eight prompts per market. Seer's response count is large but covers 20 brands in one month. AirOps publishes no methodology at all. These are vendor studies from companies selling AI visibility products, and the direction is consistent enough to trust while the precision is not.

Nobody has priced a ghost citation. It is intuitive that an unnamed citation builds no brand equity, and that intuition is probably right on a surface where most sessions end without a click. But no published work compares conversion or recall between named and unnamed citations. The cost is asserted, not measured.

A ghost citation is still a citation. It means your page was retrieved, judged useful, and linked — the hard part of earning AI citations already worked. Treating a 61.7% ghost rate as 61.7% failure inverts that. The failure is narrower: the last step, attribution, did not happen.

Seer recommends three layers, in an order that matches how much control you have over each.

Seer's own caveat on timing is the one to internalise: expect weeks to months before any of this shows up, not days. Entity strength is not a deployable change.

The practical instruction is short. Stop recording one visibility flag and start recording two, then cross-tabulate them per engine and per prompt intent. The moment you do, the question changes from "how visible are we" to "which of the four outcomes are we getting, and on which prompts" — and only the second question has a next action attached to it.

Elmo is an open-source, self-hosted AI visibility platform that runs your prompt sets across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces, recording brand mentions and cited domains as separate fields on every run alongside the competitors named next to you. Because you run it yourself and the data is yours to query, the four-quadrant cross-tab above is a GROUP BY over your own results rather than a feature request.

For the fundamentals, start with AI citations and AI share of voice, then how to track your brand in AI search. For how the sources themselves shift underneath you, see citation volatility and multi-turn AI search. For the vocabulary, see the AI search glossary.