Ghost Citations: Why AI Search Cites Your Content but Never Names Your Brand

Citation vs. Mention: Two Metrics That Are Not the Same Thing

A citation is a link. The engine references your URL as a source for a claim — a numbered footnote, a source chip under the answer, an inline superscript. A mention is your brand name appearing in the answer text that the reader actually reads, with or without a link attached.

Writesonic’s own breakdown of mentions versus citations frames them as separate events with separate causes. A citation confirms your content contributed to the answer. A mention confirms your brand travelled with it into the reader’s head. You need both, and one does not imply the other.

“Citation rate tells you that an engine found your content. Mention rate tells you whether your brand actually traveled with it.”

An independent Semrush study of the ghost citation problem — 3,981 domain appearances across 115 prompts, 14 countries, and four engines — split the outcomes cleanly. 74.9% of appearances came with a citation. Only 38.3% came with a mention. And the overlap, the appearances that had both, was just 13.2%.

Read that middle bar again. In the Semrush sample, roughly one in eight brand appearances delivered the complete outcome — a named brand the reader can remember, plus a link they can click. Everything else was partial: 61.7% cited without a name, 25.1% named without a link.

Why Two Studies Say 40% and 62% — and Why Both Are Right

If you have seen both figures circulating this month, the gap is a measurement artifact, not a contradiction. The two studies use different denominators.

• Writesonic’s ~40% is the share of citations that lacked a brand mention, measured from live tracking across roughly 16 million brand appearances in a 30-day window — a very large, naturally distributed sample.
• Semrush’s 61.7% is the share of all brand appearances that were cited but not mentioned, from a controlled prompt set built to compare engines and countries head to head.
• Engine mix drives the rest. Semrush tested ChatGPT, Gemini, AI Overviews, and AI Mode; Writesonic also included Perplexity, Grok, and Copilot, whose rates sit at both extremes of the range.

The direction is what matters, and both agree on it emphatically: most of the time you are cited, your name does not come along. Kevin Indig reached the same conclusion writing on the ghost citation problem in Growth Memo, and the framing has since been picked up across the GEO tooling space. When three independent datasets converge on a pattern this large, it is a structural property of how these systems generate text — not a quirk of one vendor’s tracker.

Why Engines Drop Your Name: Retrieval vs. Memory

An AI answer is assembled from two sources that behave very differently. There is what the model retrieves — the live pages it fetches, reads, and links. And there is what the model already knows — the brands baked into its weights during training.

Citations come from retrieval. Mentions overwhelmingly come from memory. When a model writes “the best options are Ahrefs, Semrush, and Moz,” it is reciting entities it already holds, then attaching whatever sources it happened to fetch as support. Your page can be the entire evidentiary basis for a sentence and still lose the naming slot to a brand the model recognized before it ever searched.

That is exactly what a geoSurge study published July 30, 2026 measured. Testing 66 US buyer prompts 60 times each — 3,960 responses, 13,281 fan-out searches, and 1,416 brand-level observations — the researchers found models searched for brands they already knew 3.2 times more often than unfamiliar ones. Familiar brands appeared in 55.7% of searches versus 17.4% for lesser-known companies, and 63% of brand-specific searches involved one of each model’s five most familiar names.

Put the two studies together and the mechanism is clear. Familiarity decides who gets named. Retrieval decides who gets linked. If your brand is not already resident in the model’s memory, you can win the retrieval game every single day and still ship your traffic to the incumbent whose name is on the sentence.

It also explains the engine ordering in the chart above. Perplexity and AI Mode run aggressive multi-query retrieval and cite prolifically — high citation volume, low naming density. Gemini and Copilot lean harder on model knowledge and produce more conversational, brand-naming prose with fewer links. Same content, different assembly, opposite outcome for your brand.

What This Breaks in Your AI Visibility Reporting

The practical damage from ghost citations is not the missing link — it is the missing memory. Here is what silently degrades when you track only citations:

• Downstream branded search disappears. A meaningful share of AI-influenced conversions arrive days later as branded queries or direct visits. Those only happen if the reader retained your name — which a ghost citation never gave them.
• Your competitor gets the credit for your work. When the model names a familiar incumbent and cites your page as the evidence, you have effectively funded a competitor’s recommendation.
• Reporting looks healthiest right before it fails. Citation counts can climb steadily while mention rate erodes, so the dashboard shows green until a branded-traffic decline finally surfaces in analytics a quarter later.
• Content decisions get misdirected. If informational pages generate your citations, an optimization program built on citation data will keep producing more informational content — the exact format with the worst mention rate.

The Semrush data makes that last point concrete. Informational content earned citations in 89.3% of appearances but produced brand mentions only about 18% of the time. Comparative and evaluative content produced roughly 2.4x more mentions. Informational content earns links; comparative content earns names.

There is a geographic wrinkle too. Semrush measured brand mention rates ranging from 18% to 50% depending on country, with India and Sweden at the top and Italy, Brazil, and the Netherlands at the bottom. If you report a single global mention rate across a multi-market program, you are averaging away the only signal that would tell you where to act.

How to Turn Citations Into Mentions: The Playbook

You cannot make a model name you. You can make it much harder for the model to describe your idea without saying your name. Five moves, ordered by how quickly they pay off.

1. Pair the claim to the name in the sentence

Models lift claims, not paragraphs. If your key statistic lives in a sentence that says “our research found…” three paragraphs away from any brand reference, the extracted claim arrives unbranded. Rewrite so the attribution is inside the same clause as the fact: “PikaSEO’s 2026 analysis of 4,000 SERPs found X.” It reads slightly more formal. It survives extraction.

2. Publish comparative content, not just explainers

Comparative queries — “X vs Y,” “best tools for,” “alternatives to” — force models to produce named entities, which is why they generate 2.4x the mention rate of informational pages. If your content library is 90% how-to guides, you are optimized for the citation column and nothing else. Use a Content Brief Generator to spin up comparison and evaluation briefs alongside your explainer pipeline.

3. Build proprietary, named research

A generic best-practice claim can be sourced from a hundred pages, so the model has no reason to name any of them. A finding that exists only in your dataset cannot be described without a pointer back to you. Named frameworks, annual indexes, and original studies are the most reliable mention-generators available — this article is citing Writesonic, Semrush, and geoSurge by name for exactly that reason.

4. Get onto the third-party surfaces models trust

Writesonic’s analysis of which sources influence AI answers measured roughly a 3x lift in AI visibility from presence on YouTube (3.19x), Reddit (3.17x), and analyst sources such as Gartner (3.15x), compared with relying on your own domain. Third-party mentions do double duty: they feed retrieval today and they build the familiarity that decides naming tomorrow.

5. Make the machine-readable layer say your name

Structured data is the one place you can state entity identity unambiguously. Organization and Article schema with a consistent name, sameAs, and author block gives retrieval systems an explicit brand-to-content binding rather than making them infer one. Generate valid markup with the Article Schema Generator and add answer-shaped Q&A blocks with the FAQ Schema Generator, which produces the tight claim-plus-attribution pairs models extract most cleanly.

How to Measure Citations and Mentions Separately

Fixing the reporting is faster than fixing the content, and it tells you where the content work should go. A workable measurement setup takes four steps.

On the Google side, start with the queries that actually trigger a generated answer. Our free AI Overview Analyzer shows which of your terms produce AI Overviews and where your pages surface in them, which is the natural starting sample for a mention audit. Pair it with the Helpful Content Checker on the pages that get cited most — those are the pages doing the work, and the ones worth rewriting first so the brand sits inside the extractable claims.

One caution on Search Console: its AI-surface reporting still blends AI Mode data into existing performance buckets in ways that make isolating generated-answer behavior difficult, and it has no concept of a brand mention at all. It can tell you about clicks. It cannot tell you whether anyone learned your name.

What to Expect Next

Expect mention rate to become a standard column in AI visibility tooling within the next couple of quarters. The vendors publishing this research are the same ones building the trackers, and “cited but invisible” is too sharp a sales argument to leave on the table. Treat the metric as real, but treat any single vendor’s absolute number as provisional until definitions converge.

The engine-level picture is likelier to shift than the underlying mechanism. ChatGPT has been moving toward more inline brand links, and Google has been adding source treatments to AI Mode and AI Overviews through 2026 — both changes push in the direction of more visible attribution. What will not change soon is the split between retrieved evidence and remembered entities. As long as models name what they know and cite what they fetch, brands without model-level familiarity will keep supplying uncredited evidence.

The strategic read is that AI visibility work has quietly become brand work. The teams that will win the naming slot are the ones running recognizable, repeatable, attributable programs — named research, comparative content, and a real presence on the third-party surfaces models trust. That is a slower play than technical GEO, and a much more durable one.

Tools & Resources

Frequently Asked Questions

Key Takeaways

Ghost citations are not a bug in AI search — they are what happens when a system that retrieves evidence from the open web writes prose from entities it already knows. Being cited proves your content was useful. Only being mentioned proves your brand was visible. Most AI visibility programs have been measuring the first and reporting it as the second.

The link was never the whole prize. Start by mapping where AI answers already use your content, then work through the free PikaSEO tools that make your name impossible to strip out of the claim.

Related Reading & Free Tools

More Articles