Why You Rank on Google but AI Won’t Name You: The Ghost-Citation Problem

“We rank first for half the terms that matter, and ChatGPT has never once said our name.” Some version of that sentence is now heard in partner meetings across professional services. The rankings report looks healthy. The traffic is steady. And yet when a prospective client asks an AI assistant to recommend a firm – or asks the question your best article answers – the response either names your competitors or names nobody at all.

The instinct is to assume the AI simply hasn’t found you. The research says otherwise. When Semrush and Kevin Indig measured how AI engines treat the brands they draw on, 61.7% of citations turned out to be exactly this: the engine used the domain as a source and never named the brand behind it. The result suggests the system used material from your domain, but the answer never connected that material to your brand – an omission that can arise from how the answer is constructed and summarised, or how entities are handled, not from any deliberate withholding of the name.

What is a ghost citation?

The term comes from the Ghost Citations study published by Semrush in June 2026 with Kevin Indig. The headline finding is the uncomfortable one: 61.7% of AI citations never mention the brand by name. The engine links to the domain, or draws on its content, and attributes the substance to no one. Indig’s companion analysis is blunt about what that means – being cited and being named are different achievements, and most cited brands are only managing the first.

For a professional services firm, the distinction is everything. A citation your prospect never notices builds no reputation. If an AI assistant explains inheritance tax planning using the framework your senior partner wrote, and attributes it to nobody, the prospect learns the answer and forgets the source – because the source was never shown. You did the work; the answer got the credit.

It is worth being clear about what a ghost citation is not. It is not invisibility – an invisible firm’s content never enters the answer at all. Ghost citation is the more frustrating middle state: good enough to use, not distinctive enough to name.

What the Ghost Citations study actually found

The 61.7% headline gets quoted; the detail underneath it is more useful, because it shows that “AI visibility” is not one behaviour but several – and the engines pull in opposite directions.

Four terms recur from here on, so it is worth pinning them down. Citation – the domain appears as a source link. Mention – the brand name appears in the answer text. Recommendation – the system explicitly suggests the firm. Retrieval – the system appears to have used material from the content. These are observable output categories, not transparent explanations of the system’s internal reasoning.

The engines are asymmetric – almost mirror images. ChatGPT cites brands 87% of the time but mentions them by name in only 20.7% of answers: it is generous with links and miserly with names. Gemini is the reverse – it mentions brands in 83.7% of appearances but generates a citation link only 21.4% of the time: generous with names, miserly with links. The same content, fed to two engines, produces two different kinds of partial credit. A firm optimising for “AI visibility” as a single number is optimising for neither.

The engines do not even agree with each other. In 100 of the 454 cases the study could compare across engines – 22% – the engines disagreed on whether to name the brand at all. Attribution is not a property of your content that every engine reads the same way; it is a decision each engine makes separately, and often differently.

How the question is asked changes everything. Short, conversational queries produced brand mention rates of nearly 100%, while long, structured prompts produced mention rates of just 2-3% – a 30x to 50x difference from phrasing alone. And query intent matters: comparative queries (“X vs Y”, “best firm for…”) generated 2.4x more brand mentions than informational ones. The questions where prospects are closest to choosing are the questions where engines are most willing to name names.

Geography matters more than anyone expected. In India and Sweden, 50% of AI answers included a brand name; in Italy, Brazil and the Netherlands, only 18-22% did. A firm’s likelihood of being named depends partly on which market the question is asked in.

The picture that emerges is precise: whether an AI names your firm depends on which engine is answering, how the question was phrased, what kind of question it is, and where it is asked – and in the majority of cases, across all of that variation, the brand behind the source goes unnamed. That is the ghost-citation problem, established. What the study does not establish is why any particular firm falls on the wrong side of it – which is where practice experience has to take over.

Why it happens: our working hypothesis

A necessary distinction first. The study establishes the gap – that engines routinely use sources without naming them. It does not diagnose the causes for any given site. What follows is our interpretation, not a study finding: our working hypothesis, based on the pattern we see repeatedly in professional-services sites, is that four factors commonly contribute to a firm being usable but unnameable.

• The brand is not tied to the expertise on the page. The article explains the law, the tax treatment, the regulatory position – impeccably – but nothing in the content itself asserts who is speaking or why they are qualified to. The engine can lift the substance without ever needing to lift the name, so it does.
• Missing or thin Organization and Person schema. Structured data is how you tell machines, unambiguously, that this content belongs to this firm and was written by this named expert. Without Organization and Person markup, the engine is left to infer the relationship – and inference is exactly where attribution goes missing.
• No consistent entity footprint beyond your site. Engines corroborate. A firm whose name, address and details appear consistently across directories, professional registers and industry coverage is a confirmable entity; a firm that exists only on its own domain is a URL. This matters more than many firms assume, because your own site is where the engines are looking first: BrightLocal’s July 2025 research found that “the vast majority of sources across every single LLM and industry were businesses’ own websites”. Your site is the source. Whether the engine names the firm behind it depends on how confidently it can connect the two.
• Content any firm could have written. This is the quiet killer. If your article on shareholder disputes is competent, generic and interchangeable with the top five competing articles, the engine has no reason to attribute the answer to you specifically – the information exists everywhere, so it belongs to no one. Generic content is ghost-citation bait.

We hold these as hypotheses, not laws – the engines do not publish their attribution logic, and the study’s own engine-by-engine disagreement shows there is no single rule to reverse-engineer. But notice what the four have in common: none of them is a rankings problem. That is why the frustration takes firms by surprise – every signal they have been optimising for a decade is green, and the signal that now matters was never on the dashboard.

The four stages of AI visibility

The study’s most practical lesson is that “AI visibility” needs to be split into stages, because an answer can show one outcome without the next. We frame it as four questions, in the order the engines answer them:

Each stage is a separate observable outcome, and the study shows the stages genuinely come apart: ChatGPT’s answers routinely carry a citation without a mention; Gemini’s routinely carry a mention without a citation. A ghost citation is precisely the state of clearing stage two and failing stage three – and a firm can fail at any stage while succeeding at the one before it. Most firms measuring “AI visibility” are measuring only stage four, which tells them almost nothing about where the pipeline is breaking.

This staging is also how our AI Visibility Score is built: it tracks these outcomes – whether you are cited, named and recommended – engine by engine across a fixed set of buyer-intent prompts, so you can see which stage you are losing at rather than guessing from a single aggregate number.

How to fix the ghost-citation problem

In our audits, ghost citations often point to an entity-clarity problem rather than a pure content-quality problem – and if that reading is right, the fixes follow directly. In order of priority:

1. Establish one consistent firm identity. The same firm name, the same details, everywhere you appear – your site, profiles, directories, registers, press. Every consistent mention makes you a more confirmable entity; every variant spelling makes the engine less certain who you are.
2. Put named, verified people behind the content. Named authors with biographies, credentials, and a consistent presence across the site and the wider web. An article by a verifiable expert at a verifiable firm is far harder to use anonymously than an article by “Admin”.
3. Build an independent entity footprint. Engines corroborate what your site claims against what the rest of the web says. Presence in professional registers, reputable directories and industry coverage turns the firm from a URL into a confirmable entity.
4. Write distinctive, first-hand content. Put the firm and its people into the substance, not just the byline. “Our disputes team acted on this scenario in the High Court last year” is attributable; a neutral recitation of the law is not. Original observations, real matter experience and a stated point of view are the one thing competitors cannot have written – and the claim only makes sense with the claimant attached. For regulated or confidential work, first-hand expertise does not require revealing client identities – anonymised scenarios, published decisions, methodology and professional experience all serve.
5. Implement accurate structured data. Mark up the firm as an Organization, your experts as Persons with credentials, and each article’s authorship explicitly. Schema is a machine-readable way to express the relationship between the firm, its people and its content – search and AI systems may use that information alongside other evidence.
6. Measure citation, mention and recommendation separately. Track each stage on its own, engine by engine, so you can see which one you are losing at rather than optimising a single blended number.

One caution before treating that list as a formula: no published technical rule guarantees that schema, author bios or brand references will make an AI system mention a firm. These practices improve clarity and corroboration; they do not control the final answer.

Ghost citations are one symptom of a broader pattern – the site-level habits that quietly suppress a firm’s AI visibility. We cover the full set in our pillar on the website mistakes that keep firms out of AI search, and if you want the mechanics of how engines actually parse your pages – and how differently each platform sources its answers – see what AI actually reads, platform by platform.

How to spot it happening to you

You cannot see ghost citations in your rankings report, which is precisely why most firms suffering from them do not know it. But you can run a rough diagnostic yourself, this afternoon, with nothing more than the engines and a list of questions.

• Ask the engines your buyer-intent questions. Not “what does our firm do” – the questions a prospective client actually types: “how do I contest a will in England”, “best commercial property solicitor for a retail lease dispute”, “do I need an accountant for an EIS investment”. Ask each one in ChatGPT, Gemini and Google’s AI surfaces, and note two things separately: are you linked, and are you named? The study’s engine asymmetry means you should expect different answers from different engines – that is normal, and it is exactly what you are measuring.
• Look for your themes travelling without your name. Read the answers to the questions your best content addresses. If the framework, the examples, the distinctive angles you know are yours appear in the response – and your firm does not – record it as a possible ghost-citation signal, and compare the wording with your own pages and with other likely sources before concluding the content is yours travelling unnamed. Several firms often say similar things; the signal is worth logging, not treating as proof.
• Compare branded and unbranded phrasing. Ask “how does [your firm] approach X” and then “how should a firm approach X”. If the engine describes you accurately when prompted by name but never volunteers you unprompted, it knows who you are – it just is not connecting your entity to the expertise. That gap is the ghost-citation gap in miniature.
• Vary the phrasing. The study found short, conversational queries produce brand mentions at nearly 100% while long, structured prompts drop to 2-3% – so test both. If you only ever appear in one phrasing style, your visibility is narrower than a single test would suggest.

The systematic version of this exercise is the AI Visibility Score: a fixed, tracked set of buyer-intent prompts, scored engine by engine and period over period, so the diagnostic becomes a trendline rather than an anecdote. And closing the gap it reveals – the entity, schema and content work above, done systematically – is the substance of generative engine optimisation. It is not a rebuild of your SEO; it is the attribution layer your SEO never needed until now.

FAQ