Most AI citations never say your name
Two separate things can happen to your brand inside an AI answer. The engine can use one of your pages as a source, and the engine can say your name in the text. Most of the time it does one without the other, and which one you get depends on the engine. Your analytics see neither event, only the small fraction of readers who click through afterwards. So when someone asks how the brand is doing in AI search, which number are you actually holding?
Cited and named are two different events
Semrush ran a study with Kevin Indig in June 2026 that separates them and counts each one. They logged 3,981 domain appearances for 115 prompts across 14 countries and four engines: ChatGPT, Gemini, Google AI Mode and AI Overviews. Every appearance was sorted into one of three buckets. A ghost citation is the one worth naming: the engine links your page as a source and your brand name never shows up in the answer text. Semrush puts the damage plainly. "When an AI cites your site without explicitly mentioning you, readers walk away without recognizing your brand or trusting you as the original source."
The split came out at 61.7% ghost citations, 13.2% cited and mentioned, 25.1% mentioned with no citation. Roughly six appearances in ten were work your site did anonymously. One in four was the opposite problem, a brand talked about in the answer with no link back to it. The complete result, where the engine both used your page and said your name, was the smallest of the three.
Be careful with the denominator, because a much bigger Semrush number is circulating alongside this one. The expanded AI Visibility Index covers 126 million US prompts from January to April 2026, and it is a different study with a different sample. It never uses the phrase ghost citation. Its version of the same gap is an overlap figure between the brands mentioned and the domains cited, and that overlap runs as high as 64% on AI Overviews and as low as 30% on Gemini. Two studies, seventeen days apart, pointing the same way at very different scales. The 61.7% belongs to the small one.
ChatGPT cites and does not name. Gemini names and does not cite.
The per-engine numbers are the part that should change how you report. In the ghost study, ChatGPT cited 87% of the time and mentioned brands by name 20.7% of the time. Gemini mentioned 83.7% of the time and cited 21.4%. Same behaviour, opposite direction. A brand tracking only citations will look strong on ChatGPT and near-invisible on Gemini, and the same brand tracking only mentions will read exactly backwards, with nothing about the business having changed in between.
Citation density explains half of it. The 126 million prompt index measured 15.4 sources per response on ChatGPT, 11.4 on Google AI Mode, 9.2 on AI Overviews and 3.3 on Gemini. An engine reaching for fifteen sources per answer will link plenty of pages it has no room to name in the prose. Brand density in the same index runs close to the reverse order, with 5.2 brands per response on AI Mode, 4.7 on Gemini, 3.7 on AI Overviews and 3.3 on ChatGPT. Fifteen links, three names.
The reason ChatGPT reaches so wide is the fan-out. It turns one question into a spray of smaller searches and reads what each one returns, which is how a single question ends up pulling dozens of sources. Every extra sub-search adds another cited page and no extra room in a paragraph that a person still has to read.
The direction is agreed. The size is not.
A far larger dataset lands on a different number. Writesonic looked at roughly 16 million brand appearances across seven engines over a thirty-day window and found that around 40% of citations did not name the source brand. That is four thousand appearances against sixteen million, and the bigger sample says the problem is smaller.
The engine ranking flips too. Writesonic's per-engine rates run Perplexity 52%, Google AI Mode 49%, AI Overviews 41%, ChatGPT 37%, Gemini 25%, Grok 22% and Copilot 19%. Semrush has ChatGPT at the ghost-heavy extreme. Writesonic has it second from the bottom. They cannot both be right about the ordering, so treat it as unsettled and measure your own. The caveat travels with the figures: Writesonic sells the tooling, and the sample cannot be audited from outside the company. The same caveat sits on Semrush, on Ahrefs, on us. Almost every number here came from somebody who sells a way to fix the problem it describes.
What survives both datasets is the shape. Being used as a source and being named are separate outcomes, they happen at different rates, and no engine delivers them together as the default. That holds at 40% and it holds at 62%.
More citations do not mean more mentions
The most uncomfortable number came later. Working through a six-month slice of Semrush data, Kevin Indig reported in July 2026 that the correlation between citations and brand mentions came out slightly negative at -0.229, across 1,094 US categories and more than 600,000 ChatGPT citations from January to June 2026. In the same dataset, the most cited domain in a category was also the most mentioned brand in only 20.8% of cases.
Read that carefully before you act on it. It is ChatGPT only, US categories only, and it was published as a Substack note rather than a full methodology write-up, which makes it the freshest and the thinnest-documented number here. A correlation of -0.229 is weak. What it rules out is the comfortable assumption underneath most AI-visibility reporting, that citations are a reasonable stand-in for brand presence and climbing one drags the other up with it. On this evidence they move independently at best.
The older correlation work points the same way about which one to chase. Ahrefs studied 75,000 brands in May 2025 and found branded web mentions correlate 0.664 with AI Overview visibility, against 0.218 for backlinks. That is a year-old study on a filtered sample, and correlation is not causation, but it remains the strongest correlation evidence anyone has published on the question. Mentions of your name across the web track AI visibility roughly three times better than links do.
Referral traffic is the wrong instrument
A ghost citation produces no click by definition, and a named brand with no link produces no click either. Even when the link is there, the channel is small. AI referral traffic sits at 1.08% of all website traffic on average, growing roughly 1% month on month. Reporting AI performance from analytics means reporting a one percent slice of a channel whose main effect leaves no trace in the logs at all. That is one reason brands lose AI citations for months without noticing.
Marketing teams know the instrument is wrong. In Semrush's survey alongside the index, 45% of marketing leaders said they cannot accurately measure brand visibility in AI answers, and only 9% have the tools to track all the relevant metrics across platforms. The other problem is that one engine tells you nothing about the next: 5WPR puts the overlap between ChatGPT and Perplexity at about 11% of domains, and finds 88% of cited URLs sitting on pages that do not rank in Google's top ten. The sources feeding these answers are a supply chain you mostly do not own and cannot edit, which is why watching your own site is not a measurement of it.
What you actually do
Four things, and none of them is a trick for converting a ghost citation into a named one. Nobody has demonstrated that reliably, including Semrush, whose own suggestion of adding clearer brand references to content that gets cited but not named is untested advice rather than a measured result.
1. Report the two columns separately. Citations and mentions are different events with different rates, so put them in different columns and never add them together. A single AI visibility score built by blending them will move for reasons nobody in the room can explain, because a gain in one can sit on top of a loss in the other.
2. Split both columns by engine. An 87% citation rate on ChatGPT beside a 21.4% citation rate on Gemini is the same brand having two entirely different weeks. Semrush deliberately presents its four platforms separately rather than combining them into a single weighted score, and that is the right call to copy. Track four numbers rather than one average that hides the inversion.
3. Treat a ghost citation as a content signal. A page cited without your name is a page an engine already trusts. That is useful information about which of your material is doing the work, whatever it does for recognition. Rank those pages, then decide whether the answer they feed is one you want to be attached to.
4. Measure in the markets you sell in. The ghost study ran across 14 countries. The 126 million prompt index is US only, so the industry concentration figures and the survey results in it describe one market. If you sell in three, watch three.
Mention scans are the piece most teams are missing, because the named-but-unlinked column is the one that never reaches analytics and never reaches a rank tracker either. Run one on your own brand and the split becomes a number you can put in front of a board instead of an argument about why referrals look so small.
Questions people ask about this
Sources: Semrush, the Semrush AI Visibility Index and its methodology pages, Kevin Indig and Growth Memo, PPC Land, Search Engine Land, Ahrefs, 5WPR and Analyze, drawn from publicly available reports and studies published between May 2025 and July 2026.
