# Ghost Citations: Why AI Cites Your Page but Not Your Brand

> A ghost citation is an AI answer that links your page but never names your brand. What four studies measured, our 196-citation test, and how to fix it.

- Published: 2026-11-15
- Author: Samy BEN SADOK
- Canonical: https://geotoolbox.ai/blog/ghost-citations

---

Ghost citations are the AI answers that link your page as a source and never mention your brand. Your research shapes the answer, the reader sees a tidy list of recommendations, and your name is nowhere in it.

A citation report counts that as a win. In our small test of 48 answers from ChatGPT, Gemini, Google AI Mode, and Perplexity, the answer never named the brand in 144 of the 196 citations.

## What Is a Ghost Citation?

A **ghost citation** is an AI answer that links one of your pages as a source but never says your brand's name in the text the reader sees. Your content earned a source link. The person reading the answer still walks away without knowing it came from you.

In June 2026, [Semrush and Kevin Indig measured it](https://www.semrush.com/blog/the-ghost-citations-study/) across ChatGPT, Gemini, Google AI Overviews, and Google AI Mode. Their study scored every brand appearance on two separate checks: was the domain linked as a source, and was the brand named in the answer?

Every AI answer on a topic you cover puts your domain in one of four buckets.

<figure className="not-prose my-8">
  ![Two-by-two grid of AI answer outcomes: cited and named, cited but not named (the ghost citation), named but not cited, and neither.](/blog/ghost-citations/ghost-citation-four-outcomes.png)
  <figcaption className="mt-3 text-center text-sm text-gray-500">The four outcomes when an AI answer uses a domain. A ghost citation is the top-right bucket: linked, never named.</figcaption>
</figure>

The phrase is not standardized. Four meanings are in circulation.

<table>
  <thead>
    <tr><th>Meaning</th><th>What you see in the answer</th><th>Who uses it this way</th></tr>
  </thead>
  <tbody>
    <tr><td><strong>Cited, not named</strong> (the common generative engine optimization, or GEO, meaning)</td><td>Your URL in the sources, your brand absent from the text</td><td>Semrush, Writesonic, Reach</td></tr>
    <tr><td><strong>Competitive ghost citation</strong></td><td>Your URL cited, your brand absent, a competitor named instead</td><td>Seer Interactive</td></tr>
    <tr><td><strong>Used, not shown</strong></td><td>Your content paraphrased with no visible link at all</td><td>Growth Memo's glossary</td></tr>
    <tr><td><strong>Named, not linked</strong></td><td>Your brand named, no link to your site</td><td>Promptwatch's glossary (others call this an unlinked brand mention)</td></tr>
  </tbody>
</table>

The last row is usually called an unlinked brand mention, which we cover in how to [track brand mentions in AI search](https://geotoolbox.ai/blog/track-brand-mentions-in-ai-search). The academic meaning, a fake reference invented by a language model, has its own section near the end. Here, "ghost citation" means the first row.

## How Common Are Ghost Citations?

Common, but the four public datasets give very different numbers because they measure different things.

<table>
  <thead>
    <tr><th>Study</th><th>Sample</th><th>Engines</th><th>Headline finding</th><th>What the finding is a share of</th></tr>
  </thead>
  <tbody>
    <tr><td><a href="https://www.semrush.com/blog/the-ghost-citations-study/">Semrush and Kevin Indig</a> (June 2026)</td><td>3,981 domain appearances, 115 prompts, 14 countries</td><td>4</td><td>61.7% cited, not named</td><td>All appearances, including brands named with no link</td></tr>
    <tr><td><a href="https://searchengineland.com/ghost-citation-problem-ai-483794">Writesonic, via Search Engine Land</a> (July 2026)</td><td>About 16 million brand appearances, 30 days</td><td>7</td><td>About 40% of citations unnamed</td><td>Cited appearances</td></tr>
    <tr><td><a href="https://github.com/IvanNon/ghost-citations-study-2026">Reach's open dataset</a> (June to August 2026)</td><td>22,375 answers linking a tracked brand's own site</td><td>4</td><td>43% unnamed</td><td>Answers linking the brand's site</td></tr>
    <tr><td><a href="https://www.buzzstream.com/blog/ai-mentions-vs-citations-study/">BuzzStream, Xofu data</a> (July 2026)</td><td>12,000 responses, about 200 brands</td><td>4</td><td>69.9% named, the rest unnamed</td><td>Citations of a brand's domain</td></tr>
  </tbody>
</table>

Start with the denominator. Semrush's 61.7% is a share of every appearance, and 25.1% of those appearances were brands named without any link, which can never be ghost citations. Counted against only the 74.9% of appearances that carried a link, the 61.7% cited-only share is the large majority, so on the other studies' basis Semrush's rate is even higher and the gap between studies widens.

Another split is whose domains get counted. Semrush counted every domain that appeared, and reference sites are cited far more often than they are named. Reach and BuzzStream counted brands they track, and Writesonic describes its sample as thousands of brands. Every source here sells something: Semrush ran its study on its own AI Visibility Toolkit, Writesonic's CEO co-wrote the analysis for Semrush-owned Search Engine Land, Reach and BuzzStream sell tools, Seer is an agency, and we sell a tracker. Read it all as vendor research.

The engines also disagree with each other. In Writesonic's data, Perplexity left the brand unnamed in 52% of its cited appearances, Copilot in 19%.

<figure className="not-prose my-8">
  ![Bar chart of ghost citation rates by AI engine in Writesonic's data: Perplexity 52%, Google AI Mode 49%, Google AI Overviews 41%, ChatGPT 37%, Gemini 25%, Grok 22%, Microsoft Copilot 19%.](/blog/ghost-citations/ghost-citation-rate-by-engine-writesonic.png)
  <figcaption className="mt-3 text-center text-sm text-gray-500">Share of cited appearances that left the brand unnamed, by engine. Writesonic data via Search Engine Land, July 2026.</figcaption>
</figure>

[Reach's data](https://github.com/IvanNon/ghost-citations-study-2026) also puts Perplexity highest, at 54%, but orders the rest differently: ChatGPT 43%, AI Overviews 40%, AI Mode 36%. ChatGPT is where the studies split hardest: in [Semrush's data](https://www.semrush.com/blog/the-ghost-citations-study/) it cited 87% of its appearances but named only 20.7%, while [BuzzStream](https://www.buzzstream.com/blog/ai-mentions-vs-citations-study/) measured GPT-5.4-mini naming 92.7% of the brands it cited.

Seer Interactive uses the strictest definition, a citation where your brand is absent and a competitor is named in the same answer. In [its client data](https://www.seerinteractive.com/insights/llm-ghost-citations-why-your-content-is-working-and-your-brand-isnt), awareness-stage prompts had the highest rate by funnel stage (5.0%), industrial services the lowest by industry (0.3%), and two hospitality and travel brands differed by more than 20 percentage points. Reach adds that [87% of its ghost answers](https://github.com/IvanNon/ghost-citations-study-2026) named another brand instead.

## What We Found Testing 48 AI Answers

To see the pattern first-hand, we ran a small test at geotoolbox. On November 15, 2026, we sent 12 everyday questions to ChatGPT, Gemini, and Google AI Mode (logged out, from the US) and to Perplexity through its Sonar API: four informational questions, four "best" questions, and four how-to questions across CRM software, VPNs, personal investing, and running shoes.

For each of the 48 answers, we listed every domain the engine cited and checked whether the answer text named that domain's brand. One cited domain in one answer is one citation. Source chips that only display a domain name counted as citations, not mentions; a credit line such as "Source: TechRadar" counted as a mention; and ChatGPT's product link-outs that were not among its search results did not count at all. We collected the answers through DataForSEO, with ChatGPT's web search forced on and Gemini answering with its Flash-Lite model. Seven answers cited nothing, six of them from Gemini.

**In 144 of the 196 citations, the answer never named the brand.** That is 73%, far above the studies that count only tracked brands and below Semrush once Semrush is counted on linked appearances only.

<figure className="not-prose my-8">
  ![Stacked bars of citations where the brand was named or not named in our 48-answer test: ChatGPT 20 named and 21 not named, Google AI Mode 22 and 47, Perplexity 7 and 43, Gemini 3 and 33.](/blog/ghost-citations/ghost-citation-test-by-engine.png)
  <figcaption className="mt-3 text-center text-sm text-gray-500">Our test, November 15, 2026: 12 prompts per engine, US, one run. A citation is one cited domain in one answer.</figcaption>
</figure>

The engine split is lopsided. Gemini named 3 of its 36 citations, from only six answers, the reverse of Writesonic's tracked-brand Gemini figure, and Perplexity 7 of 50, while ChatGPT named 20 of its 41. By question type, "best" questions had the highest unnamed share (68 of 88), ahead of informational (46 of 63) and how-to questions (30 of 45).

The more useful pattern is which domains got named. Brands that were the answer got named: Fidelity was cited 7 times and named in 6 of those answers. The sites that supplied the comparison got cited and ignored: NerdWallet was cited 7 times and named once, RunRepeat 5 times and named once. That matches [Semrush's finding](https://www.semrush.com/blog/the-ghost-citations-study/) that aggregators and reference sites get cited while strong brands get named.

Treat these numbers as a snapshot: one run, small samples, every cited domain counted.

## Why AI Engines Cite You Without Naming You

Being retrieved and being recommended are separate decisions, and your page can pass the first while your brand fails the second.

[Seer Interactive's leading hypothesis](https://www.seerinteractive.com/insights/llm-ghost-citations-why-your-content-is-working-and-your-brand-isnt) is that citations are often added after the answer is drafted. The model picks which brands to name from what it learned in training, then retrieval finds pages to support the answer, so your guide becomes evidence for a list of brands chosen without you.

Seer's supporting signal: a brand's citation rate was 53.1% in responses that named it and 10.6% in responses that did not. Seer calls this strong behavioral evidence rather than proven architecture.

Answer design matters too. Engines that attach many links per answer cite more domains than they name: [BuzzStream measured](https://www.buzzstream.com/blog/ai-mentions-vs-citations-study/) about 33 cited URLs per AI Mode response against 3.8 for GPT-5.4-mini. Engines also read more than they credit, as with [how ChatGPT cites sources](https://geotoolbox.ai/blog/chatgpt-citations), and a 2025 [attribution study](https://arxiv.org/abs/2508.00838) of about 14,000 conversations found Perplexity's Sonar visited about 10 relevant pages per query but cited three to four.

The question type also changes the outcome. [Semrush's data](https://www.semrush.com/blog/the-ghost-citations-study/) by query type:

<table>
  <thead>
    <tr><th>Query type</th><th>Citation rate</th><th>Brand mention rate</th></tr>
  </thead>
  <tbody>
    <tr><td><strong>Informational</strong> ("what is", "how does")</td><td>89.3%</td><td>18%</td></tr>
    <tr><td><strong>Comparative</strong> ("best", "vs")</td><td>Not reported</td><td>43.3%</td></tr>
    <tr><td><strong>How-to</strong></td><td>Not reported</td><td>42.8%</td></tr>
    <tr><td><strong>Commercial</strong> (pricing, buying)</td><td>84.4%</td><td>35.6%</td></tr>
  </tbody>
</table>

These are shares of all appearances. Comparison answers name more brands, but a larger share of their citations may go unnamed: BuzzStream's list prompts named the brand in only [36.5% of citations](https://www.buzzstream.com/blog/ai-mentions-vs-citations-study/). Phrasing matters even more: short conversational prompts drew brand mentions almost every time, long, structured ones in just 2% to 3% of answers. Reference sites take the worst of it: medium.com was cited 16 times in Semrush's sample and never named.

## Do Ghost Citations Still Send Clicks?

Possibly. A ghost citation is still a link a reader can click, but no study here measured its clicks.

The closest data is indirect and comes from Google's AI Overviews. On informational queries in [Seer Interactive's 2026 click-through analysis](https://www.seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update), being cited in an AI Overview delivered about 120% more organic clicks per impression than not being cited, and still trailed queries with no AI Overview by 38%. Seer notes that one account drives much of the not-cited baseline, says it cannot claim causation, and did not split cited pages by whether the brand was named.

None of these studies measured the brand cost; the Writesonic analysis says it does not establish whether ghost citations affect clicks, brand recall, trust, or conversions.

So the answer depends on your business model. For a publisher, a citation that brings a click is a fair result. If you sell a product, a citation under a list of competitors may do more for their visibility than for yours.

## How to Measure Your Ghost Citation Rate

A spreadsheet is enough. Ask what the studies asked: were you linked, and were you named?

<figure className="not-prose my-8">
  ![Five-step workflow for measuring a ghost citation rate: build a prompt set, run it on each engine several times, log the full answer and every cited URL, sort each answer into four buckets, then calculate the rate per engine every month.](/blog/ghost-citations/measure-ghost-citation-rate-workflow.png)
  <figcaption className="mt-3 text-center text-sm text-gray-500">Run the same prompts, keep the full answers, and classify every appearance of your domain.</figcaption>
</figure>

Run the same prompt set on each engine, log every cited URL, and sort each answer into one of the four buckets:

1. **Build a prompt set.** Mix short and long prompts, and "what is" questions with "best" and "vs" questions.
2. **Run each prompt on every engine, more than once.** Answers vary from run to run.
3. **Save the full answer text and every cited URL,** not just a yes or no for your domain.
4. **Classify each answer for your brand:** cited and named, cited only, named only, or neither.
5. **Calculate the rate per engine:** cited-only appearances divided by all appearances where your domain was cited. Report the counts with the percentage, monthly.

Fix your definitions before you start: which names count (brand, product names, misspellings) and whether a source chip showing only your domain counts as a mention.

For a stricter version, borrow Seer's check: of your cited-only appearances, how many named a competitor instead?

Google's Search Console now has [generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports), rolled out to all websites as of August 31, 2026. They show how often your URLs appeared in AI Overviews and AI Mode, not whether the answer cited the URL as a source or named your brand, so they cannot show ghost citations.

This rate sits alongside the [core AI search visibility metrics](https://geotoolbox.ai/blog/ai-search-visibility-metrics) rather than replacing them. To automate this kind of AI citation tracking, the geotoolbox [AI visibility tracker](https://geotoolbox.ai/features/ai-visibility-tracker) records whether each answer cited you, mentioned you, or left you out, prompt by prompt.

## How to Turn a Ghost Citation Into a Brand Mention

No study has found a reliable way to make an engine name you every time. The first two levers rest on observational data; the rest are practitioner recommendations.

**Put your brand in the page title.** In [Reach's dataset](https://github.com/IvanNon/ghost-citations-study-2026), cited pages without the brand in their title were ghosted 42.9% of the time, against 12.2% for pages that had it, and within the same brands the reduction was about 60%. Reach flags the confound itself: pages with the brand in the title are often homepages, product pages, or "Brand vs X" pages, which are about the brand anyway.

**Get named in comparisons; publishing your own is not enough.** Comparative prompts name more brands, as Semrush's query-type table shows, but your own comparison pages are no shortcut: in Reach's data, brands' own listicle and "vs" pages were ghosted 50% of the time against 34% for their other pages, and "best" questions had the highest unnamed share in our test. Being named in other sites' comparisons is the bet worth testing.

**Put your brand inside the claim.** If a statistic sits on your page without your name in the same sentence, the engine can lift the fact and drop the name. Write "Acme's 2026 survey found" rather than "a recent survey found", and give original research a name people can repeat. Volume alone does not help: [Seer warns](https://www.seerinteractive.com/insights/llm-ghost-citations-why-your-content-is-working-and-your-brand-isnt) that adding more content will not fix a ghost citation rate and may make it worse.

**Strengthen how the web describes you.** Seer's list: consistent naming, Organization schema with sameAs links, Wikidata entries, author schema tied to the organization, and FAQ answers that name the brand. The aim is to connect your name to your category. Seer argues that third-party coverage naming you in recommendation contexts is where models learn which brands belong on a list.

**Don't expect quick results.** [Seer had a client add](https://www.seerinteractive.com/insights/llm-ghost-citations-why-your-content-is-working-and-your-brand-isnt) explicit brand language to a post that had been cited more than 100 times with zero mentions. Twenty-nine days after the change went live on February 20, 2026, mentions on that page were still at zero.

For the offsite signals, geotoolbox's [Citation Interceptor](https://geotoolbox.ai/features/citation-interceptor) lists the Reddit threads, forums, editorial pages, and competitor pages AI engines cite for your topics where your brand is absent.

## Ghost Citations in Academic Writing

In research and legal writing, a ghost citation means something else entirely: a reference to a paper, book, or court case that does not exist, produced by a language model that predicts what a plausible citation looks like. These are also called ghost references or fake citations.

The [GhostCite study](https://arxiv.org/abs/2602.06718) tested 13 language models on citation tasks and found they hallucinated citations at rates from 14.23% to 94.93%. It also checked 2.2 million citations from 56,381 published AI, machine learning, and security papers and found invalid citations in 1.07% of papers, with an 80.9% increase in 2025.

Checking a reference is quick:

1. Search the exact title in quotation marks on Google Scholar or the publisher's site
2. Resolve the DOI and confirm it points to the same title, authors, and year
3. Open the source and confirm it says what the citation claims

Our guide to [AI hallucinations](https://geotoolbox.ai/blog/ai-hallucinations) covers the brand-facing side of invented facts.

## Frequently Asked Questions

### Is it okay to use ChatGPT for citations?

Use it to find sources, not to write your reference list. When ChatGPT searches the web, it links real pages you can open and check. When it answers from memory, it can produce references that look correct and do not exist, so verify the title, authors, and DOI before you cite a source.

## Sources

- Why 62% of AI citations don't lead to brand mentions [Study] - Semrush and Kevin Indig, June 9, 2026 - `semrush.com/blog/the-ghost-citations-study`
- Ghost citations: Why AI search cites your content, not your brand - Search Engine Land (Writesonic data), July 29, 2026 - `searchengineland.com/ghost-citation-problem-ai-483794`
- Ghost Citations Study 2026 (open dataset) - Ivan Nonveiller, Reach, 2026 - `github.com/IvanNon/ghost-citations-study-2026`
- How AI Mentions and Cites Your Brand (New Study) - BuzzStream, updated July 14, 2026 - `buzzstream.com/blog/ai-mentions-vs-citations-study`
- LLM Ghost Citations: Why Your Content Is Working and Your Brand Isn't - Seer Interactive, March 2026 - `seerinteractive.com/insights/llm-ghost-citations-why-your-content-is-working-and-your-brand-isnt`
- AIO Impact on Google CTR: 2026 Update - Seer Interactive, April 2026 - `seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update`
- The Attribution Crisis in LLM Search Results - Strauss, Yang, O'Reilly et al., arXiv, June 27, 2025 - `arxiv.org/abs/2508.00838`
- Introducing Search Generative AI performance reports in Search Console - Google Search Central, June 3, 2026 - `developers.google.com/search/blog/2026/06/gen-ai-performance-reports`
- GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models - Xu et al., arXiv 2602.06718, v2 May 14, 2026 - `arxiv.org/abs/2602.06718`
- Ghost citation (glossary) - Growth Memo, Kevin Indig, August 21, 2026 - `library.growth-memo.com/p/ghost-citation`
- Ghost Citation (glossary) - Promptwatch - `promptwatch.com/glossary/ghost-citation`
