Ghost Citations: Why AI Picks Your Rival

Your page gets cited. Your competitor gets recommended. That is a ghost citation. It is now one of the most common outcomes in AI search.

In research circles, "ghost citation" also means a fabricated reference. This article covers the marketing meaning: a real citation that never names its source brand.

A Semrush study run with Kevin Indig found 61.7% of AI citations were ghost citations. Reach found that 87% of ghost answers named other brands instead. Most teams read this as a content problem. It is a sequencing problem.

Mike King offered the cleanest mental model for AI search at Profound's Zero Click NY in 2026. AI search is a raffle. The system fans one prompt out into many sub-queries. Every sub-query you appear in is another ticket. More tickets, better odds.

The raffle model is right about citations. It is wrong about recommendations. Ghost citations are the proof: brands with plenty of tickets keep losing the draw.

The reason is structural. AI search runs three gates, and each gate awards a different prize. In AI search, model memory decides which brands get shortlisted. Retrieval decides which pages get cited. Judgment decides which brand gets picked. Most GEO programs only work on the middle gate.

Related readings

What are ghost citations? Three prizes, three inputs

A ghost citation occurs when an LLM cites your URL but names a competitor in the answer. Your content does the work. Someone else gets the credit.

Seer Interactive analysed 541,213 LLM responses across 20 brands. When a brand was mentioned in the answer, its content citation rate hit 53.1%. When it was not mentioned, that rate fell to 10.6%. The two signals travel together. Seer's data shows correlation, not which one causes the other.

That split makes sense once you separate the prizes.

Prize	What it is	What earns it	How fast it moves
Citation	Your URL in the sources	Retrieval coverage across fan-out sub-queries	Days to weeks
Mention	Your brand named in the answer	Category depth and third-party corroboration	Weeks to months
Recommendation	Your brand named as the choice	A shortlist slot plus evidence that meets the model's criterion	Months

The prizes are related. They are not interchangeable. Optimising for one rarely buys the others.

Retrieval is not citation either. Resoneo's ChatGPT study found arXiv pulled more than 2,600 times. It was cited 10 times.

Query intent shifts the odds. In the Semrush study, informational queries cited sources 89.3% of the time. They named brands only 18% of the time. Comparative queries named brands 43.3% of the time. Buying questions are where mentions happen, and where ghosts cost the most.

This is why citation dashboards mislead. A rising citation count can sit alongside a flat recommendation rate for months. The team celebrates. The pipeline does not move.

Which kind of ghost citation do you have?

Not every ghost citation has the same cause. The fix depends on which kind you have.

Attribution ghosts happen at the page level. The engine trusts your page as a source but cannot tie its content to your brand. Reach studied 71,275 AI answers that cited a brand's own page. When the brand was missing from the page title, 42.9% of those citations left it unnamed. With the brand in the title, the rate fell to 12.2%.

Within each brand, Reach found brand-titled pages were ghosted 60.3% less often. That held in 17 of 20 brands. Repeating the brand on the page did not help. Pages with 10 or more brand mentions were ghosted 45.9% of the time.

Reach is clear that this is observational data. Brand-titled pages are often homepages and product pages. Still, a title change is cheap, so test it.

Shortlist ghosts happen at the memory level. The engine names a competitor because your brand was never on its list. In Reach's engine-level sample, 87% of ghost answers named other brands. A title fix will not solve this kind. Only the memory and judgment work below will.

Most teams have both kinds. Clear the attribution ghosts first, because that fix takes weeks. What remains is the shortlist problem, and that is where recommendations are won.

Diagnose the failed gate

Four patterns tell you which gate failed. Each one points to a different fix.

Cited often, rarely named in fan-out queries: a memory problem. Fix corroboration.
Named in fan-out queries, rarely recommended: a judgment problem. Fix criterion evidence.
Ghosted on your own pages: an attribution problem. Fix page titles first.
Rarely cited anywhere: a retrieval problem. Only here is more coverage the answer.

The measurement section below shows how to get each number.

This is where the raffle model breaks. The model does not search a neutral field. It searches the brands it already believes in.

Suganthan Mohanadasan reverse engineered ChatGPT's retrieval across 57 conversations in August 2026. ChatGPT writes brand names into its first search query before it fetches anything. It then runs one site: probe per named brand, aimed at that brand's own website. The fan-out works through a list the model already holds.

The gap in outcomes is stark. Brands named in ChatGPT's own query reached the answer 68.9% of the time. In ChatGPT, fetched-only brands reached the answer 2.1% of the time. Of 3,554 retrieved pages, 110 earned a citation.

Suganthan is careful about the limits. The sample is one account, weighted toward software queries. He calls the decimals directional. The ordering is the point. In Suganthan's sample, named brands reached ChatGPT answers 33x more often than fetched ones.

geoSurge found the same direction on Gemini. Its team ran 66 buyer prompts 60 times each, producing 13,281 fan-out queries. Brands in Gemini's top-10 memory for a category appeared in fan-out queries 55.7% of the time. Brands outside that memory appeared 17.4% of the time. geoSurge found Gemini fan-outs named remembered brands 3.2x more often.

Most fan-out queries were generic. Only 31% named a specific brand. But 63% of those brand-led queries named one of the model's top-5 remembered brands. When the model goes looking for brands, it walks its memory list.

Seer's Gemini 3 research points the same way. In Gemini 3, 26.4% of fan-out queries named a brand. Seer frames it carefully: a fan-out mention does not guarantee the answer, but it lowers the barrier.

The challenger exception

Memory is not the only door. In geoSurge's payments example, Gemini searched Stripe, PayPal and Square from memory. It also searched Lemon Squeezy, a brand it never recalled. Web presence pulled an unremembered brand into the search.

So challengers can still get in. They just enter at retrieval, and retrieval mostly pays out in citations. That is exactly how shortlist ghosts are born.

geoSurge names the confound too. Famous brands are both remembered and searched more. The study measures association, not cause. Treat it as direction, then test it in your own category.

What does query fan-out coverage actually buy?

Breadth is not useless. It buys a specific prize.

Semrush and Kevin Indig tracked 1,094 US categories in ChatGPT from January to June 2026. They analysed 283,215 citation observations and 76,493 brand-mention observations. The two outcomes behaved differently.

In ChatGPT, citation share rose with prompt coverage. Appearing in even one of five prompt variants was associated with some citation gain. Appearing in all five gained more. There was no spread-thin penalty for citations.

Mentions needed depth. Brands present in only one of five prompts had lower mention share. That penalty faded once a brand appeared in three of five.

Relatedness mattered as well. Citations named the brand 46% of the time in closely related categories. In the least related categories, that fell to 18%.

Industry changes the maths. In finance and real estate, wider coverage held citation leverage without hurting share. In legal and healthcare, mention share trended negative even at full coverage.

This is King's raffle, measured. More tickets buy more citations. King conceded the limit himself in the same talk. Pages ranking deep in the SERP rarely make the AI Overview. Even the raffle has a price of entry.

So run the raffle for citations. Just do not expect it to produce recommendations.

Why do named brands still lose? The judgment gate

Being named and cited still does not mean winning. The model has to decide that your evidence passes the test it applies to the query.

Yunhee Choi of Artience showed this failure at Search SEOul. A car brand's evidence stressed premium finishes and elevated comfort. The model defined value as what came in the base trim. The brand's facts were correct. Its criterion was wrong. It was excluded.

Nothing in a citation tracker catches that failure. The brand was retrieved. It may even have been cited. It lost because its evidence answered a question the model was not asking.

Write evidence to be remembered with its criterion

This is the move most GEO programs miss. Memory does not just store your name. It stores the contexts your name keeps appearing in.

Suppose reviews, comparisons and press repeatedly pair your brand with one buying criterion. The model learns that pairing. Later it recalls your brand together with the test you have to pass.

That pairing shapes the fan-out itself. Take an asset tracking platform that third parties keep describing as working without signal. Sub-queries like "asset tracking that works offline" start to carry its name. The brand stops competing inside the query space. It becomes part of how the query space is built.

So start from the criteria the model already applies, not the positioning you wish it applied. Run your buying prompts across engines. Read the reasoning where models expose it. Extract the tests they use to separate options. Then pick the criteria you can genuinely win and build evidence for those only.

A criterion you invent is one the model will ignore. A criterion the model already applies is a lever.

How to harvest the criteria

Criterion harvesting takes a day, not a quarter. Four steps cover it.

Build the buying prompt set. Write 20 to 40 prompts a real buyer would ask. Include comparisons, alternatives and use-case questions.
Run them across engines. ChatGPT, Gemini, Perplexity and AI Mode each apply different tests. Run each prompt several times to see which criteria repeat.
Extract the tests. Note every phrase the model uses to include or exclude an option. Exclusion statements are the most useful signal.
Filter hard. Keep a criterion only if it passes three tests. It moves the buyer's choice. You can prove it. The model does not yet credit you with it.

That last filter matters most. A criterion the model already credits you with is won. A criterion you cannot prove is a liability.

How should B2B SaaS teams fix ghost citations?

The gates run in sequence, so the work should too.

Resolve the entity. The model must know what you are and which category you sit in. Keep your naming identical across your site, Wikidata, Crunchbase, G2 and LinkedIn. Put the brand name in the titles of pages engines already cite. Schema supports this as a disambiguation and corroboration layer. It is not a visibility lever on its own. Wikidata and other knowledge graphs give models one consistent record to anchor mentions to. Consistent identifiers help models consolidate scattered mentions into one remembered entity.
Build corroboration that feeds memory. In SaaS, Aleyda Solis found 82.3% of AI citations come from external sites. Social and community platforms alone carried 45.7%. Your own site corroborates. Third parties earn the shortlist. Reviews are also how memory learns your criterion. A G2 review saying "works offline" pairs your name with that test. Enough of them, and the model recalls the pairing. Third-party coverage across adjacent categories also builds the mention depth Semrush measured.
Match evidence to the criterion. Publish atomic, verifiable facts that satisfy the tests buying prompts trigger. Then get third parties to repeat those facts in their own words.
Cover the routes. Only now does query fan-out coverage pay off. Match the format each sub-query routes to: comparison pages, review platforms, video and documentation. Fix comparison pages first. Reach found "best X" pages had the highest ghost rate, 49.5% without the brand in the title.

What this means for the budget

If most of your GEO budget funds owned content, it mostly buys retrieval. Retrieval is the cheapest prize. The shortlist is earned off-site.

Shift the weighting toward the gates that decide recommendations. Fund digital PR, review programs and community presence that repeat your criterion claims. Keep owned content focused on atomic evidence those third parties can quote.

Most teams run this backwards. They start at step 4 because it looks like familiar SEO. They end up with a citation report full of ghosts.

This sequence mirrors how we work at Exalt Growth. Entity-First covers steps 1 and 2. Proof of Importance covers what makes a chunk citable at step 4. The judgment gate sits between them, and it is where most recommendation losses happen.

How do you measure ghost citations at each gate?

One blended visibility score hides which gate failed. Track four numbers instead.

Metric	Gate	How to get it
Named-in-fan-out rate	Memory	Share of fan-out queries that name your brand, from tools that expose fan-outs
Citation rate	Retrieval	Share of tracked answers that cite your URLs
Ghost citation rate	Retrieval vs mention	Share of answers citing you that do not name you
Recommendation rate	Judgment	Share of buying prompts where you are the named pick

Split every number by engine. Ghost rates vary too much to blend.

Engine	Reach: own pages without the brand in the title	Writesonic: all brand appearances
ChatGPT	54.2%	37%
Perplexity	52.3%	52%
Google AI Overviews	37.4%	41%
Google AI Mode	33.9%	49%
Gemini	Not measured	25%
Microsoft Copilot	Not measured	19%

The two studies measure different things. Compare engines within a study, not across them. The Writesonic analysis covered about 16 million brand appearances over 30 days. About 40% of those appearances used a brand's content without naming the brand.

Separate pulled from cited as well. Retrieval is not citation. Citation is not recommendation.

Read the four numbers together, then match the pattern to the gate diagnostic above.

Run the raffle for citations. Earn the shortlist for everything else.

FAQs

What is a ghost citation?

A ghost citation occurs when an AI answer cites your URL but never names your brand. Often it names a competitor instead. Your content informs the answer while another brand receives the recommendation.

ChatGPT tends to write brand names into its search queries before it retrieves anything. Brands named that way reached the answer 68.9% of the time in one study. Brands that were only fetched reached it 2.1% of the time. Your page can be retrieved and cited without your brand ever being shortlisted.

How do I fix ghost citations?

Start with the pages AI engines already cite. Put your brand in their titles, beginning with comparison pages. If engines still name competitors, the problem is memory. Build third-party corroboration around the criteria buyers care about.

Which AI engine has the most ghost citations?

ChatGPT and Perplexity ghost most in current studies. Reach found ChatGPT ghosted 54.2% of citations to pages without the brand in the title. Writesonic's data put Perplexity highest at 52%. Track each engine separately.

Does query fan-out decide which brands get recommended?

Fan-out shapes the candidate set, but memory shapes fan-out. geoSurge found remembered brands appeared in Gemini fan-out queries at 3.2 times the rate of other brands. Fan-out tends to confirm a shortlist the model already holds.

Is a ghost citation the same as a hallucinated citation?

No. A hallucinated citation points to a source that does not exist. A ghost citation points to a real page, but the answer never names that page's brand. One is a fabrication problem. The other is an attribution and recommendation problem.

What is the difference between an AI citation and an AI recommendation?

A citation links your URL as a source. A recommendation names your brand as the answer to the buyer's question. Citations respond to retrieval coverage. Recommendations need memory, corroboration and evidence that meets the model's criterion.

What evaluation criteria do LLMs apply when recommending software?

They vary by category and prompt. Models apply tests such as offline capability, base-tier inclusions or pricing transparency. To find them, run buying prompts and extract the criteria from answers and reasoning traces.

How long does it take to change parametric memory?

Plan in months, not days. Retrieval changes can show within days. Memory shifts depend on new third-party signals entering training data and model update cycles. Seer's tests of brand-claim changes showed no effect at 29 days.

How do I measure my ghost citation rate?

Take the answers that cite your URLs and count how many fail to name your brand. That share is your ghost citation rate. Track it per engine, next to your named-in-fan-out rate and recommendation rate.