A marketing manager checks whether AI is citing the company blog, finds three ChatGPT answers linking to it, and reports back that AI visibility is working. Then somebody reads the actual answers. The links are there. The company name is not mentioned once.
This has a name now. Kevin Indig called it the ghost citation, and a study he ran with Semrush put a number on how common it is. Across 3,981 domain appearances from 115 prompts, run in 14 countries across ChatGPT, Gemini, Google AI Overviews and Google AI Mode, 61.7% of appearances were citations with no brand mention in the answer. Only 13.2% got both. Put differently: 74.9% of appearances included a citation, but only 38.3% included the brand name.
Your content can be doing all the work while your brand stays completely invisible to the person reading.
The engines are not doing the same job
The split between engines is severe enough that a single visibility score is close to meaningless.
- Gemini names the brand in 83.7% of appearances but generates a citation link only 21.4% of the time. It answers from what it already knows about you.
- ChatGPT does the reverse: cites 87% of the time, names the brand in 20.7%. It reads like a paper with footnotes.
- Google AI Overviews sit between the two and lean towards citing.
In the study’s data there was almost no overlap between the brands ChatGPT cited and the brands Gemini named for the same prompt. On 100 of 454 prompt and domain combinations, the engines disagreed on whether to name the brand at all. So if you are tracking one engine and reporting it as “AI visibility”, you are reporting one behaviour of one system.
The content type doing the damage is the one you produce most
This is the finding that should change what a small marketing team publishes next month.
Informational content, the “what is”, “how does”, “explained” format that most business blogs are built from, earned an 89.3% citation rate and an 18% mention rate. Comparative content, the “best”, “versus”, “recommended” format, produced a 43.3% mention rate, roughly 2.4 times more brand mentions. How-to content sat at 42.8%. Commercial queries reached 35.6% mentions against an 84.4% citation rate.
The logic is uncomfortable but obvious once you see it. Informational content is raw material. The model reads it, absorbs the fact, writes the fact in its own words, and links the source out of politeness. Comparative content forces the model to name things, because a comparison without names is not a comparison.
So the agency that spent eighteen months building a definitive glossary of industry terms has produced the single most citable and least brand-visible asset available. Meanwhile the competitor with six honest “X versus Y” pages, including the cases where they lose, gets named.
Why the model will not repeat your claim
There is a second layer under the format question, and Forbes framed it well this week: every brand claim an AI repeats without independent backing is a claim the AI is vouching for itself. Systems tuned against Google’s quality rater guidelines weight first-hand experience heavily, and a product page describing your own jacket as the warmest available has an obvious incentive problem. A forum thread describing a Colorado winter in that jacket does not.
This is where most GEO advice goes vague, so here is the concrete version. If your only evidence for a claim lives on your own domain, expect the model to use the surrounding information and skip the claim. Corroboration that actually shifts this looks like: a named customer willing to be quoted with specifics, a third-party review platform with enough volume to be more than a testimonial page, a trade publication write-up you did not pay for, forum and community threads where customers describe outcomes in their own words, and comparison content that names competitors honestly enough to be credible.
None of that is fast. That is the point. There is no schema markup fix for a claim nobody else has ever repeated.
What to actually do
A sequence that fits inside a normal fortnight:
- Pull the ten prompts that matter most to your business. Run each one in ChatGPT and Gemini separately, and record two columns: cited, and named.
- Find the pages that are cited but never named. These are your ghost pages, and they are usually your best informational content.
- For the top three, add a comparison or evaluation section that names real alternatives, including where you are not the right choice.
- Pick one claim you make constantly and find or build one piece of third-party evidence for it that is not a testimonial on your own site.
- Re-run the ten prompts in six weeks. Compare the named column, not the cited column.
Query phrasing matters too, and it cuts against most keyword habits. Short conversational prompts produced brand mention rates near 100% in the study, while long structured prompts on the same topic produced 2% to 3%. Test the short version of your question, because that is how people actually ask.
What can go wrong here
The obvious risk is over-correcting. If you decide informational content is worthless and stop making it, you lose the citations that build the model’s familiarity with your domain in the first place. Citations and mentions come from different places: citations reflect domain authority and original work, mentions reflect brand familiarity and positioning. You need both, and killing one to chase the other is how sites lose ground in both.
The second risk is measuring this too often. Answers vary run to run. A single check on a single day tells you almost nothing, which is exactly the trap the sample-of-one approach falls into. Six-week intervals, ten fixed prompts, two engines minimum.
And this may not be worth your time at all. If you sell into a market where buyers do not use assistants for discovery, if your pipeline is referral and relationship driven, or if you are a publisher whose business model runs on the citation link rather than the name, then citation rate is the correct metric and the ghost citation is not your problem. Check who your buyers actually are before rebuilding a content plan around this.
The one thing worth taking away: stop treating a citation as proof of visibility. It is proof your content was useful to a machine. Whether it was useful to your business is a separate question, and it has a separate column in the spreadsheet.
The full dataset is in the ghost citations study, and the trust argument is set out in this piece on AI search and branded content. If you want the other half of the picture, what these systems can and cannot read on your site, we covered that in how AI builds a B2B shortlist before anyone calls you.
Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.