There is a failure mode in AI search that most dashboards never surface because they only count one thing. You check an AI-visibility tool, see that your pages are getting cited, and conclude you are winning — but open the actual answers and your brand name is nowhere in the text. The engine quoted your fact, linked your URL in the sources, and then named a competitor as the recommendation, or named no one at all. That is a ghost citation: cited as evidence, invisible as a brand. It is not a rare edge case. A 2026 Semrush analysis run with Kevin Indig's Growth Memo, covering 3,981 domain appearances across 115 prompts, 14 countries, and four AI search engines, found that 61.7% of appearances were cited without the brand being named — against just 13.2% that were both cited and named. The engines also split hard: ChatGPT cited a domain in 87% of its appearances but named a brand in only 20.7%, behaving like an academic paper that footnotes everything and credits no one in the prose, while Gemini did almost the reverse. This guide is the practitioner's read on that gap: what a citation and a mention actually are and why they are produced by two different steps of the system, what the data shows and how far to trust it, why collapsing both into one 'AI visibility score' hides the exact problem you need to fix, the four distinct diagnoses the gap resolves into, and the honest limits — because there is no documented rule for when a model decides to say your name.
There is a quiet failure mode in AI search that most dashboards are structurally unable to show you, because they count one thing and call it visibility. You open an AI-tracking tool, see a healthy citation count — your pages are getting pulled into answers — and conclude you are winning the channel. Then you read the actual answers a customer would read, and your brand name is not in them. The engine quoted your statistic, listed your URL down in the sources, and in the sentence a human actually reads it named a competitor as the thing to go use, or it named nobody at all. You supplied the evidence and someone else got the introduction.
That is a ghost citation: cited as a source, invisible as a brand. The term was popularized in 2026 by Kevin Indig's Growth Memo off a Semrush dataset, and the reason it caught on is that it names a gap practitioners kept hitting without a word for it. A citation and a mention feel like the same success and are not — they are two different outcomes produced by two different parts of the system, and in AI search they come apart far more often than they line up. This guide is about that gap specifically: what each term really means, how wide the divergence is in the data, the mechanism that drives it, why one blended score hides it, the four distinct problems it resolves into, and the honest limits of chasing it. It sits next to but apart from the two-lever maintenance view in AI citations, brand mentions, and content refresh and the per-asset scoreboard in content measurement in AI search; this page is narrower — it is about the gap between being quoted and being named, and what to do about it.
Start with clean definitions, because the whole problem is that the two words get used interchangeably and they describe different things. A citation is a source: the engine surfaces your URL as the evidence behind a claim, typically as a link in a sources list under or beside the answer. It tells you which page the model trusted enough to justify a statement. A mention is recognition: your brand name appears in the answer's actual sentences, the part a reader reads. It tells you which brand the model decided to put in front of the person asking. One is attribution plumbing; the other is the on-stage introduction.
The reason the distinction matters commercially is that they do different jobs. A mention is what builds awareness and gets a brand into someone's consideration set — it is the version that can send a branded search, a direct visit, or a 'I saw you in ChatGPT' later. A citation can send a click if the reader expands the sources, but on a ghost citation the reader never learns your name, so the brand-building value is close to zero even though your content did the work. You can be mentioned without being cited (the model names you from what it already knows, with no live source) and cited without being mentioned (your page supplied the fact, a competitor got the name). Because either can happen without the other, a tool that reports only one of them is reporting on a different question than the one you care about — a point argued at length in the metrics-level view in AI visibility measurement.
The gap is measurable and it is large. In the 2026 Semrush analysis run with Growth Memo — 3,981 domain appearances across 115 prompts, 14 countries, and four AI search experiences (ChatGPT, Google AI Overviews, Gemini, and Google AI Mode) — appearances split into three buckets: 61.7% were cited without the brand being named (the ghost citations), 25.1% were mentioned with no citation, and only 13.2% were both. Put the other way: across all appearances, 74.9% carried a citation but only 38.3% carried a brand mention. For most brands, most of the time, getting pulled into an answer does not mean getting named in it.
The engines are not the same, and the differences are the actionable part. ChatGPT cited a domain in 87% of its appearances but named a brand in only 20.7% — the widest gap in the set. It behaves like a footnoted academic paper: generous with sources, sparing with names in the prose. Gemini ran almost the reverse, naming brands in 83.7% of appearances while citing a source only 21.4% of the time. So the same content that earns a ChatGPT ghost citation might earn a Gemini mention with no link, and a strategy that optimizes for one engine's behavior can be invisible on the other's terms. This is why a per-engine read beats a blended one, and why the engine your audience actually uses changes what 'winning' even looks like.
Trust these numbers as direction, not as constants. The figures come from vendors that sell AI-visibility tracking, from a single study window on a system that changes how it retrieves and phrases weekly, and 'citation' and 'mention' are counted slightly differently by different tools — which is why other 2026 studies report ChatGPT leaning the opposite way on mentions, a contradiction that mostly reflects differing prompt sets and definitions rather than a real disagreement about the world. The durable takeaway survives the noise: citations and mentions diverge, often dramatically, and the direction of the gap differs by engine. The exact percentage is a snapshot; the structural fact is not.
The gap is not a glitch. It falls directly out of how an answer is assembled, in two steps that make two different decisions, and seeing the two steps is what turns 'why won't it say my name' from a mystery into a fixable diagnosis.
The first step finds and ranks the evidence. Given a query, the system retrieves pages it considers trustworthy and relevant and leans on them to ground the answer and reduce the chance of making something up. Your page earns its place here on the strength of the content itself — a clean, concrete, extractable fact that directly answers the question, from a domain the system trusts. That is the citation, and it is largely the game that classic extractable-content advice optimizes for: be the page a model can lift a confident sentence from. For the mechanics of winning this step, generative engine optimization and AI-answer citations covers the extractable-page discipline in full.
The second step writes the prose, and this is where your name lives or dies. The language model composes the answer and, sentence by sentence, decides which entities are salient enough to name — which brand is the recommendation, which is the example, which is left as an anonymous 'one tool' or dropped entirely. That decision keys off how strongly and consistently your brand is associated with the topic across everything the model has absorbed, not off the single page that got retrieved. So the model can lift your fact (retrieval worked) and still name an established competitor as the answer (generation does not see your brand as the salient entity for that claim), or name no one because the query has no clear owner. Entity recognition failed even though retrieval succeeded. This is the same mechanism behind the uncomfortable finding in AI brand citations that the source an AI quotes about your brand is increasingly not you — your fact survives, your name does not.
Three things most reliably produce a ghost citation out of that second step. Your content answers the question so completely that the brand is incidental — the model extracts the information and leaves the company behind. Your off-site footprint is thin or inconsistent — you are not named across the review sites, comparison articles, forums, and coverage that teach a model which brand owns a topic, so you have no entity salience to draw on. Or your entity itself is ambiguous — you describe the company differently across your site, your profiles, and third-party listings, and ambiguity is exactly what makes a model hesitate to put a name in a sentence it has to stand behind.
The most common measurement mistake is to blend mentions and citations into a single AI-visibility number, because a single number collapses outcomes that have opposite causes and opposite fixes. A brand with great citations and no mentions and a brand with great mentions and no citations can post the identical blended score while having completely different problems — and the score tells you nothing about which one you are. Semrush's own 2026 visibility index found that, on average, fewer than half of the top-100 brands on a category's mention list also appeared on the matching citation list, so for most brands the two rankings genuinely disagree and averaging them erases the signal.
Tracked separately, the gap resolves into four distinct diagnoses, each pointing at a different lever. A recommendation problem is when the model does not name you enough — you are cited but ghosted, an entity-salience and off-site-presence issue. An attribution problem is the inverse: the model names you but credits someone else's page as the source, so a competitor's content is carrying the evidence about you. An ownership problem is when third-party domains carry most of the evidence that shapes how the model talks about you, so you do not control the inputs. And a platform problem is when one engine surfaces you well while another barely sees you — the ChatGPT-versus-Gemini split above, live in your own data. The practical setup is a stable prompt panel run on a schedule across the major engines, broken out by prompt cluster, reporting mention share, owned-citation share, and third-party citation share as separate lines. The argument for splitting the dashboard this way is made well in the industry writing on why a single blended score hides the exact problem you need to fix.
Because the gap has four shapes, the fix depends on which one your data shows — the diagnosis is the whole point of measuring them apart. For a recommendation problem (cited, not named), the work is entity salience: make your brand the thing the generation step associates with the topic by being named consistently across the off-site surfaces models read, and by making your brand integral to the claim rather than incidental to it, so the fact cannot be cleanly lifted without the name attached. For an attribution problem (named, cited to someone else), the work is owning the evidence — publish the definitive, extractable version of the fact on your own domain so the citation lands on you instead of the comparison site currently holding it.
For an ownership problem (third parties carry the evidence), the work is presence: get your brand accurately represented across the review sites, forums, comparison content, and social surfaces that feed the model, because you cannot control what you are not present in. For a platform problem (strong on one engine, invisible on another), the work is coverage — meet each engine on its terms, since the ChatGPT-leans-citation, Gemini-leans-mention split means a single content shape will not win both. Underneath all four is the same unglamorous requirement: a consistent brand entity, stated the same way everywhere, associated with your topic densely enough and across enough surfaces that the generation step has something to reach for. That is the bridge from 'our page got quoted' to 'we got named,' and it is closer to the displace-named-competitors framing in AI brand visibility as share of model than to any single on-page tweak.
Be clear about what this gap does and does not let you control. There is no documented rule for when a model decides to name a brand — OpenAI and the others publish no naming mechanism, so there is no dial to turn, only conditions to make favorable. You are shaping probabilities over a non-deterministic system, which means any single answer is a sample: the same prompt can name you today and ghost you tomorrow, and a one-off screenshot proves nothing. Measurement has to be a pattern across repeated runs, not a verdict from one reading.
Two more cautions. The data behind the whole ghost-citation conversation is vendor-produced, single-window, and defined inconsistently from tool to tool, so treat the percentages as evidence of a direction rather than coefficients to optimize to. And closing the gap does not override the prerequisite that still governs everything upstream of it: if your content is not genuinely good enough to be cited and your brand not genuinely credible on the topic, no amount of entity tidying manufactures a mention — you cannot be named as the authority on something you have no authority on. The ghost citation is a real and specific problem with specific fixes, but it is a problem you earn the right to have by first being citable at all.
The mechanism above says a ghost citation is a second-step failure: retrieval keeps your page, generation drops your name because your brand is not salient and consistent enough across the surfaces a model reads. Closing it is therefore not a reporting task — it is a production one, and that is the specific thing Kompozy is built for. It is a content generation and multi-platform publishing engine, not a tracker, so it will not tell you your ghost-citation rate (pair it with a prompt-tracking tool for that). Its job is upstream of the measurement: to manufacture the consistent, brand-forward, multi-surface presence that gives the generation step a reason to name you instead of lifting your fact anonymously.
The root cause this guide keeps returning to — an ambiguous entity and a thin off-site footprint — is precisely what a single, enforced brand identity run across many surfaces fixes. A written Persona Brief fixes exactly how your brand is described and keeps that description identical on every asset Kompozy produces, so you stop feeding models the inconsistent self-descriptions that make an entity ambiguous. A face-locked persona pool and HyperFrames keep the look and voice the same across every video and graphic, so the brand a model sees on a carousel, a short, a blog post, and a newsletter is unmistakably one entity. Consistency is not a branding nicety here; it is the input entity salience is built from, and it is the thing a lean team cannot maintain by hand across dozens of pieces a week.
The other half of the root cause is density — being associated with your topic across enough surfaces that the generation step treats you as the owner of the claim, not an incidental source. Kompozy turns one idea into 18 output formats and fans them across the eight social platforms plus blog and email, so a single topic becomes a blog article that earns the owned citation and a spread of short video, carousels, images, and posts that build the off-site mentions the model learns your name from — the citation track and the mention track fed by one engine rather than two workstreams. Autopilot keeps that running on a cadence, and a per-post review gate means the volume never ships off-brand, which matters because an inconsistent entity shipped at scale makes the ghost-citation problem worse, not better. Kompozy does not decide what is true about your brand or supply the authority that earns the citation in the first place — those stay yours. What it removes is the production ceiling that keeps a brand quoted-but-unnamed: the reason you were a ghost citation was rarely a bad page, it was a brand too thin and too inconsistent across surfaces for the model to confidently name, and that is a volume-and-consistency problem a generation engine is built to solve.
A ghost citation is when an AI engine uses a page from your domain as a source — linking it in the answer's citations — but never names your brand anywhere in the answer text. Your content supplied the fact; the reader just never learns it came from you. It is distinct from a brand mention, where the engine names you in the prose, and the two frequently come apart: a 2026 Semrush study with Growth Memo found 61.7% of observed appearances were cited without the brand being named.
A citation is a source — the engine shows your URL as evidence behind a claim, usually as a link in a sources list. A mention is recognition — the engine names your brand in the answer's actual sentences. They answer different questions: a citation tells you which page the model trusted enough to justify a statement, a mention tells you which brand the model chose to put in front of the reader. You can have either without the other, which is why tracking only one misleads.
Very often. In a 2026 Semrush analysis run with Kevin Indig's Growth Memo across 3,981 domain appearances, ChatGPT cited a domain in 87% of its appearances but named a brand in only 20.7% — the widest citation-to-mention gap of the engines studied. ChatGPT behaves like a footnoted paper: it links sources generously but keeps brand names out of the prose. Gemini did close to the opposite, naming brands in 83.7% of appearances while citing a source only 21.4% of the time. Treat these as directional, not fixed.
Because retrieval and naming are two different steps. The retrieval step pulls your page because it is trustworthy evidence for the fact — that earns the citation. The generation step then writes the prose and decides which brands are salient enough to name, and if your brand is not strongly and consistently associated with the topic across the web, the model lifts your fact and names an established competitor, or names no one. A weak, inconsistent, or thinly-referenced brand entity gets quoted but not credited.
Do not use a single blended 'AI visibility' number — it collapses outcomes that need different fixes. Track at least three things per engine and per prompt cluster: your brand mention share (named in the answer), your owned-citation share (your pages linked), and third-party citation share (who else's pages carry the evidence about you). Run a stable prompt panel across ChatGPT, Gemini, and the Google AI surfaces on a schedule rather than taking one-off screenshots, because AI answers are non-deterministic and a single reading is a sample, not a verdict.
A ghost citation is when an AI engine cites your page as a source but never names your brand in the answer text. Citations and mentions are produced by different steps — retrieval decides what to link, generation decides whom to name — so they diverge sharply. A 2026 Semrush analysis with Growth Memo found 61.7% of appearances were cited without the brand named; ChatGPT cited 87% of the time but named a brand in only 20.7%. Measure them separately, because each gap has a different fix.
Get started → · ← All guides · Compare Kompozy vs other tools