Ghost citations in AI search expose a brand visibility gap

A citation is not the same thing as brand visibility. In Writesonic's analysis of a recent 30-day window of AI answer data across thousands of brands and roughly 16 million brand appearances, about 40% of AI citations did not name the source brand in the generated answer. On Perplexity, that gap reached 52%.

That matters because AI search often satisfies the user before a click. If the model borrows your data, links your page, and still leaves your name out, your content helped build the answer but your brand did not necessarily travel with it. For generative engine optimization, or GEO, teams, that means citations and mentions are different outcomes and should be measured separately.

What is a ghost citation, exactly?

A ghost citation is when an AI answer links to your page as a source but does not name your brand in the answer text. The engine acknowledges the page in the evidence layer while stripping the company from the visible summary.

A simple example makes the problem obvious. In one version of an answer, the model names Writesonic as the source of the finding. In another, it says only that an analysis found the result, even if the citation panel points to the same page. The reader gets the fact either way, but only one version carries the brand.

This is why AI visibility starts before the prompt and ends with citations. In AI interfaces, the answer itself is the impression. If your brand is not in that impression, the value of the citation changes. You may still earn referral traffic from attentive users, but you lose direct brand recall from everyone who reads the answer and moves on.

That distinction sounds subtle until you look at how people actually use AI search. Most users read the answer first, maybe glance at a source list, and rarely investigate every link. If the answer gives them what they need, the journey often ends there. A page can influence the outcome without the brand receiving obvious credit.

How big is the gap across AI engines?

The gap is not uniform. The same citation count can produce very different levels of brand recognition depending on the engine.

• Perplexity: 52% ghost citation rate
• Google AI Mode: 49%
• Google AI Overviews: 41%
• ChatGPT: 37%
• Gemini: 25%
• Grok: 22%
• Microsoft Copilot: 19%

One of the most useful takeaways is the split between what you could call citers and namers. Perplexity and Google's answer engines appear more willing to link out, but they are also more likely to omit the brand name from the answer. Gemini and Microsoft Copilot behave more like namers: they include brands more often in the text, but cite pages less frequently. ChatGPT and Grok sit between those two behaviors.

Put differently, 100 cited appearances on Perplexity do not mean the same thing as 100 cited appearances on Copilot. Based on this dataset, a brand would be missing from the answer in about 52 of those Perplexity appearances and about 19 of the Copilot ones. If your reporting flattens both into one citation score, it is hiding the real visibility outcome.

This is the first trap for marketing teams. A dashboard can show growth in source citations and still tell the wrong story about awareness. Two brands can post similar citation totals while one of them becomes memorable in the answer and the other remains invisible unless the user opens a footnote.

The study also needs to be read with restraint. It is a snapshot of behavior during one recent 30-day measurement window, percentages were rounded, and it does not prove the downstream effect on clicks, recall, trust, or conversions. But it is large enough to make one point hard to ignore: engine choice shapes recognition, not just traffic.

Why does this distort most AI visibility reporting?

Because many teams still treat source presence and brand presence as if they were the same KPI. They are not.

A citation tells you that an engine used your content. A mention tells you that the user actually saw your brand. The strongest outcome is getting both at once. A mention without a citation can increase awareness but may not drive referral clicks. A citation without a mention can validate your content while doing little for brand memory.

That is why any serious GEO program needs a measurement model that separates four states: cited and mentioned, cited but not mentioned, mentioned but not cited, and neither cited nor mentioned. If you want a practical framework for that, BotRank's guide on how to measure your AI visibility is a useful starting point.

The bigger strategic issue is budget allocation. A team may think its AI search program is working because source counts are rising. In reality, it may be feeding answer engines with useful material while a competitor gets the branded credit in the visible answer. That is not failure at the content level. It is a measurement failure.

There is also a reporting problem inside organizations. SEO teams often own traffic and crawlability, PR teams own reputation, and brand teams own recall. Ghost citations sit between all three. If nobody tracks them explicitly, they disappear into blended reporting and nobody gets a clean view of what changed.

Why do ghost citations happen?

The honest answer is that nobody has a complete causal model yet. What we have are plausible explanations and clear engine-level patterns.

One likely reason is that some engines treat the citation itself as sufficient attribution. They absorb the source material, rewrite it in their own voice, and leave the brand name behind in the footnotes or sources panel. From the model's perspective, the answer is still supported. From the marketer's perspective, the attribution is weakened.

Content structure may also matter. When a key statistic lives several paragraphs away from the company name that produced it, the claim is easier to anonymize during synthesis. When the brand name is built directly into the sentence, the bond between insight and source is harder to break. This is still a hypothesis, not a proven rule, but it is a smart one to test.

There is also a brand entity layer. If a model has weak confidence in how your company should be named, categorized, or connected to a topic, it may default to neutral phrasing even while citing your page. That is one reason strong entity consistency, clear page structure, and repeated attribution across the web still matter.

This is where a lot of teams overreact in the wrong direction. They assume ghost citations must mean the content is bad. That is not what this dataset shows. The issue may live in the engine's synthesis behavior, the page's attribution pattern, the brand's entity footprint, or a mix of all three. A blunt rewrite of the whole article is rarely the first smart move.

BotRank's Take

The most important shift here is operational, not philosophical. Ghost citations mean AI visibility has to be measured at the answer level, not inferred from source logs alone. If one model frequently cites you but strips your brand, while another names you but rarely links to you, the right fix is different in each case.

This is exactly where BotRank's AI Visibility feature matters. It lets teams run the same prompts across multiple LLMs, compare model-specific outcomes over time, and separate citations from visible mentions instead of collapsing them into one score. In the context of ghost citations, that difference is everything. It shows whether your content is merely feeding an answer or actually carrying your brand into the answer. That is a more useful signal for strategy, reporting, and prioritization than a single blended visibility number.

What should SEO and GEO teams do next?

The next move is not panic. It is better instrumentation and tighter attribution.

1. Measure mentions and citations separately

Start with the four-outcome model: cited and mentioned, cited but not mentioned, mentioned but not cited, and neither. Then break that down by engine. Perplexity, ChatGPT, AI Overviews, Gemini, Grok, and Copilot do not produce the same attribution behavior, so an aggregate score can hide the problem.

This is also where prompt selection matters. Track commercial questions, category comparisons, and high-intent informational prompts separately. A brand may be well cited on educational queries and invisible on buying queries. If you do not segment by prompt type, you can end up optimizing the wrong pages for the wrong outcomes.

2. Create research, datasets, and frameworks that carry your name

Generic advice is easy for AI systems to compress into anonymous language. Named research is harder to detach from its creator. A benchmark, an index, a framework, or a methodology with your brand built into the label creates a stronger attribution bond than a loose sentence about our analysis.

This works especially well when the asset is open, crawlable, and easy to quote. A gated PDF can hold valuable research, but it is a weak format for AI retrieval and citation. A public page with a clear methodology, summary bullets, and structured sections is much easier for models to extract. BotRank's post on why 90% of brands are still invisible in AI search makes the larger point well: strong traditional visibility does not guarantee AI mention or citation visibility.

The off-site effect matters too. When third-party coverage repeats the name of your study or framework, the association between brand and insight gets reinforced beyond your own domain. That gives models more chances to learn the relationship instead of rediscovering it from one page alone.

3. Put attribution inside the claim, not near it

If you publish original data, do not bury your brand name in the intro and the actual finding halfway down the page. Pair them. A sentence that explicitly ties the brand to the claim is stronger than one that assumes the reader will infer ownership from the page context.

That principle applies beyond studies. Product pages, comparison pages, expert commentary, and FAQs all benefit from explicit attribution when you want the brand to travel with the idea. If you want to inspect which pages are actually being used and whether they present your brand clearly, BotRank's Source Analysis gives teams a way to review the pages and citations sitting behind LLM answers.

4. Audit the answers, not just the URLs

When a page is cited but the brand is absent, review the answer itself. Which sentence was extracted? Were competitor brands named? Did the model paraphrase a proprietary term into something generic? Does the same omission repeat across related prompts?

This is where model-by-model testing becomes mandatory. As BotRank has argued in its analysis of ChatGPT Search citing fewer domains, AI visibility is not one stable environment. Retrieval and citation behavior can shift by model, product surface, and answer style. Diagnosis has to happen at the prompt and answer level, not just in a spreadsheet of URLs.

A useful example is a pricing or alternatives query. A model may cite your comparison page, omit your brand name, and still mention a competitor in the visible answer because that competitor is more strongly anchored as an entity. Looking only at the citation would miss the competitive loss.

5. Turn insights into a repeatable backlog

Ghost citations are not solved by one rewrite. They need ongoing testing. That means turning findings into specific tasks: rename a study, restructure an answer block, add explicit attribution, strengthen off-site coverage, or improve the sections most often cited.

BotRank's Recommendations feature is useful here because it translates diagnosis into concrete GEO actions instead of leaving teams with a pile of observations. That is important because the real bottleneck in AI search is rarely awareness of the issue. It is follow-through.

The practical goal is simple: increase the share of answers where your brand is both cited and named. That is the closest thing to a high-quality AI visibility win in this dataset.

FAQ

Does a citation still matter if the brand is not named?

Yes. It shows the engine found your page useful enough to support an answer. But it is weaker than a citation plus a mention because the reader can consume the answer without noticing the brand.

Which AI engines look most vulnerable to ghost citations in this dataset?

Perplexity showed the highest ghost citation rate at 52%. Google AI Mode followed at 49%, then Google AI Overviews at 41% and ChatGPT at 37%.

Are ghost citations caused by bad content?

Not necessarily. The study does not establish causation. Likely factors include how each engine rewrites sources, how closely a claim is tied to the brand name on the page, and how clearly the brand exists as an entity across the web.

What is the best KPI for this problem?

The best KPI is not raw citation count. Track citations and visible brand mentions separately, then watch how often they occur together for the prompts and engines that matter to revenue.

If you want to know whether AI is borrowing your content or actually carrying your brand, start by auditing prompts across engines, labeling ghost citations explicitly, and treating recognition as a first-class GEO metric. That is how invisible influence turns into visible brand demand.