Here is a thing I have had to explain to three buyers in the last month, and each one went quiet when the number landed. You pay for the content. You pay for the schema. The AI pulls it up as a source. Then it recommends somebody else by name in the same answer. The link is yours, the credit is not.
I had this filed under “I think I have seen this once or twice” until April, when Kevin Indig put a dataset behind it and gave it a name. He called them ghost citations. Semrush then ran the full study with him in June, 3,981 appearances across four engines, and the number was 61.7%. Nearly two out of every three times an AI engine cites a brand’s content, it does not say the brand’s name.
If you are a marketing lead reading this on your phone, the part you want is simple. The visibility problem you are being pitched a tool to fix is not the visibility problem you have. You are not invisible to AI. You are visible to it and inaudible through it. Those are two different jobs, and almost everyone I talk to is still paying for the first one.
What follows is Indig’s argument in his own words, Seer Interactive’s parallel study, Charles Nicholls at SimplicityDX pushing back on the obvious fix, and what any of that means inside a real WordPress or WooCommerce estate on a Monday morning.
Being Cited and Being Named Are Not the Same Transaction
Indig’s framing, from his joint write-up with Semrush on 9 June 2026, is the sentence that reorders the whole conversation: “There’s almost no overlap between which brands ChatGPT cites and which ones Gemini names for the same prompt.” Read that twice. He is not describing noise. He is describing two separate currencies.
A citation is a link next to a hedged answer. A mention is a name inside a recommendation. The second one is the one that matters commercially, because that is the one a shopper, a buyer or a procurement committee reads and acts on. Indig’s second line on the gap is the one every marketing lead I have shown this to has read twice as well: “The brand name carries on its own.” Once an engine knows a brand exists as an entity, it will name it in the sentence it writes, with or without a source. If the engine does not know, no amount of citation hygiene will change that.
How Big the Gap Actually Is
The Semrush and Growth Memo study, written up by Margarita Loktionova and Christine Skopec with data run by Indig, logged 3,981 brand appearances across 115 prompts, 14 countries and four engines (ChatGPT, Google AI Overviews, Gemini, Google AI Mode). The headline split: 74.9% of appearances carried a citation, 38.3% carried a brand mention, and only 13.2% carried both. The 61.7% that are citations with no mention are the ghost citations. The 25.1% that are mentions with no citation are the inverse ghost, your name without the credit.
The engine-by-engine picture turns the “two currencies” line operational. Gemini mentions brands in 83.7% of appearances and cites them in 21.4%. ChatGPT cites in 87% and mentions in 20.7%. If your dashboard watches one of those two numbers and reports it as “AI visibility”, you are reporting a quarter of the picture, and the quarter is almost opposite between the two surfaces.
Seer Interactive ran a parallel exercise at bigger scale and published their methodology on 10 August 2026: roughly half a million LLM responses across 20 brands. Their category finding is the one worth lifting out, because it is the one that puts a lower-bound on how much category you are losing. In Hospitality and Travel, Seer measured citation-to-mention gaps of up to twenty percentage points between equally established brands. One name surfaced in recommendation contexts, the other did not. The content budget on both sides looks the same from the outside. The outcome does not.
There’s almost no overlap between which brands ChatGPT cites and which ones Gemini names for the same prompt.
Kevin Indig, Growth Advisor, Growth Memo and Semrush study, 9 June 2026
The Actual Lever Is Entity Recognition, Not Citation Hygiene
Both studies land on the same diagnosis and most practitioners I know have not caught up to it yet. Ghost citations are not a crawling problem or a schema problem. They are a brand entity recognition problem. The engine understands that a page says a thing. It does not understand that the page is written by the brand it should be recommending. That is a different layer of the stack, and the fixes are different.
Seer’s recommendation list from the same August post reads like three jobs for three different teams. Make your brand name inseparable from your key claims (editorial). Build AI-readable signals that connect your brand to your category expertise, through schema.org, Wikidata and Wikipedia (information architecture). Earn third-party mentions that recommend your brand by name (PR, communications, partnerships). None of those is the SEO team’s job alone, and that is one honest reason the ghost citation problem has lingered for so many organisations. There is no single owner for it.
Charles Nicholls, CEO of SimplicityDX and Chair of DX Academy, wrote the sharpest counter-position on the obvious fix in a Forbes Communications Council piece on 10 August 2026. His argument against the Reddit-seeding reflex (where brands pay to appear in third-party user discussions because AI engines over-index on Reddit) is worth quoting in full: “Brands chasing Reddit placement are often optimizing for the symptom instead of the cause.” His root-cause framing is better still: “Brands should treat trust as something to be demonstrated by others, not asserted by themselves.” That is the right posture. The engines are not confused about what your brand says. They are starved of independent corroboration that it is true.
Prompt Length Changes Everything, and Your Buyers Prompt Short
The single most actionable finding in the Semrush study, and the one I almost missed on first read, is the prompt-length behaviour. Short, conversational prompts (the “best accounting software for a small agency” register) hit nearly 100% brand mention rates. Long, structured prompts hit 2 to 3%. The study calls it a 30x to 50x difference, and the prompts in the dataset ranged from 16 to 98 characters with an average around 60.
That is a usage pattern, not an optimisation lever. The real-world prompts your buyers type are short. The exhaustive, structured prompts an “AI SEO audit tool” simulates on your behalf are long. If your reporting runs against synthesised long prompts, you are measuring a scenario where the engines behave worst for mentions and best for citations, which is also the scenario where your tool vendor’s dashboard looks most defensible. The honest audit uses the prompts the buyer actually types.
The country numbers in the same study are the second shoe to drop. India and Sweden hit 50% mention rates. Italy, Brazil and the Netherlands sit at 18 to 22% mention rates despite 82 to 94% citation rates. If you run a multi-market store, the gap you are instrumenting at HQ is one market’s gap, and it is not the one you need to budget against.
Third-Party Corroboration Beats Yet More Owned Content
Nicholls points at the real traffic signal in the Forbes piece, citing Peec AI’s analysis of roughly 30 million cited sources: Reddit, YouTube and LinkedIn dominate AI citations because they contain real user discussion. His framing is the clean one: “AI models aren’t loyal to Reddit, but they are starved of alternatives.” That is a very different reading than “we need to post on Reddit”. It says the asset class the engines want is first-person corroboration, and Reddit is only the biggest and messiest supply of that asset class.
For a mid-market brand, this reframes two budgets I see fought over every quarter. The content budget is defensible if what it is producing can be independently repeated by somebody who does not work for you. Case studies with named customers who will say the thing on camera count. Review platform integrations that are honest about the mixed reviews count. A blog post that an AI engine can quote only on your authority does not count the same way. The engines are weighting the independent signal.
The Lever Nobody Has Cracked
In the interest of being straight with you: there is no published, controlled intervention study where a brand changed one of these variables and the ghost citation rate moved in a reproducible way. Seer recommends tracking the ghost citation rate as a monthly KPI, segmented by platform and funnel stage. That is a measurement, not a lever. It will tell you whether you are getting better or worse. It will not tell you which change you made moved it.
I would rather tell you that than let an agency pitch you a six-figure “GEO retainer” against a KPI that nobody has shown is causally movable. The honest version of the brief is: you are instrumenting your side of a black box, you are weighting the levers the research supports, and you are prepared to let a quarter pass before you draw conclusions.
What This Looks Like Inside a Real WordPress or WooCommerce Install
Econsultancy stops here. We do not, because the next question in the room is “fine, how do we actually ship this”. Five pieces, in the order I would stage them inside an estate we build and run.
One: an entity disambiguation layer. schema.org Organization markup with a complete sameAs array pointing at the brand’s Wikidata item, its LinkedIn page, its Crunchbase page, its Companies House listing where that applies, and its canonical social profiles. Managed as a WordPress option, injected site-wide via a theme hook or a thin mu-plugin, so there is one source of truth. Half a sprint if the Wikidata item already exists, one to two sprints if it has to be created and cross-referenced, which it often does for younger brands.
Two: brand-name density audit at the claim level. A WordPress query that pulls every post and page and reports, per claim-bearing paragraph, whether the brand name appears within the paragraph or only in the header and footer. Not a readability score; a proximity score. Content with the brand name near the key claim is content an engine can quote without having to fish it out. One sprint to build the audit, then two to four weeks of editorial work to close the gaps, run as a recurring Query Monitor job thereafter.
Three: third-party review structure that an engine can read. schema.org Review and Rating markup aggregated from G2, Trustpilot, Capterra or whichever platforms your category uses, with sameAs pointing at your profile on each. On WooCommerce, native reviews are a start; the gap is making the third-party ones machine-readable on your own domain under a clear review schema, rather than hidden behind a JavaScript widget the AI crawler cannot execute. One to two sprints, depending on how many sources you need to reconcile.
Four: a ghost citation KPI dashboard. A GA4 custom dimension for AI-referred sessions keyed by referrer domain and user agent, plus a quarterly manual audit of a representative prompt set across ChatGPT, Perplexity, Gemini and Google AI Mode. The quarterly audit is unglamorous; it is also the only part that measures mentions, because the referrer logs only measure citations that someone clicked. Two to three days of engineering plus a standing half-day per quarter from an analyst who speaks the brand’s language.
Five: a content model change for brand-near-claim. An Advanced Custom Fields or block pattern that prompts editors to place the brand name within the claim paragraph by design, not by remembering. The content model does the enforcement so the next six editors do not have to re-learn the lesson. One week to build, then it is invisible.
Rough total: eight to ten weeks of build, plus a quarter of measurement before anyone has a defensible read on whether it is working. Any shorter timeline is somebody selling you a dashboard, not a result.
What I Would Tell a Buyer Deciding Between a Tool and a Team
The tools in this space mostly measure one of the two currencies, usually the citation side, because citations are a URL and URLs are easy to parse. Useful but not sufficient. The mention side needs human reading of what the engine wrote, which is why the honest audits still involve someone sitting with a prompt list on a Tuesday afternoon. If somebody quotes you a figure for an “AI visibility score” computed entirely from log data, ask them which of Indig’s two numbers it maps to. If they do not have an answer, you have your answer.
If you want the five pieces above shipped against your WordPress or WooCommerce estate without the quarterly audit becoming the next thing nobody owns, that is the kind of work we take on. The honest conversation is one call, and we come to it with the measurement framing before we come to it with a quote.
Last modified: November 1, 2026
