SEO Galaxy is now GetMentioned. The platform stays the same. Want a demo? Book your slot here →
A brand appears in an AI answer with no source attached. Four different cases carry the same name, and only three of them concern brand visibility. What the studies measure and how the gap can be closed.
David Hahn · August 26, 2026 · 34 min read
The short version
A ghost citation is a source reference in an AI answer that comes without the brand name in the answer text. The page is listed as a source link. The reader still does not learn who the information came from.
The term marks a gap between two events that reports routinely add into one figure. A citation is one event. A mention is the other.
For a brand that difference decides what a mention is worth. A mention nobody can attach to a sender still reaches the reader. It cannot be traced afterwards.
This article sorts the figures that exist, separates four cases that run under the same name, and shows what follows for measuring brand visibility. It is written for anyone handed a number on AI visibility who wants to know what it counts.
A ghost citation happens when an AI system links a page as a source and leaves out the brand name behind that page. The brand supplies the content. It gains no visibility from doing so.
For the reader the answer then looks like information without a sender. There is a statement, below it a row of source links, and the brand name appears at no point.
The term comes from AI visibility and is young. It was coined in 2026 and quantified for the first time in the same year.
A citation is a clickable source reference below or inside an AI answer. It points at an address. It says nothing about whether a name appears in the text.
A mention is the brand name in the answer text itself. It sits in the paragraph the user actually reads. It needs no link.
Both events can occur together. They can also occur on their own. That separation disappears the moment a report folds both under the word visibility.
The two events also arise at different points in the answering process. The citation arises when the sources are selected. The brand name arises when the answer text is written.
That explains why the two values move independently of each other. A system can pull a page and then avoid every proper noun while writing. Another system can place a name from model knowledge and supply no source for it.
The only study so far that counts both events separately defines them briefly. A citation is recorded when "The domain appeared as a source link in the response."
For the mention the definition reads "The brand name appeared in the AI answer."
Both sentences stand in the analysis by Semrush and Kevin Indig of 9 June 2026. They are the yardstick every figure below is measured against.
Semrush credits Kevin Indig with coining it. The wording there is "Kevin Indig coined the term ghost citation to describe this gap."
Indig published the finding on 20 April 2026 under the title "The ghost citation problem" in his newsletter Growth Memo, drawing on the same dataset. The full text there sits behind a paywall. The figures are openly available in the Semrush post.
Authorship for the AI visibility sense is therefore documented by the study's data partner and not by an independent body. We report it as it stands there.
In research the term means something else. There it refers to a reference with no existing work behind it. The more common English expression for that is ghost reference.
Both meanings use the same picture. Something stands in the reference that carries nothing in the result.
They still describe different things. In AI visibility the name is missing from the text. In research the work behind the reference is missing.
A query for the term returns hits from the marketing press and hits from academic journals side by side. Both use the same two words for two different problems.
This article deals mainly with the AI visibility sense. The research sense gets a section of its own, so it does not slide in unnoticed while you read.
The term is fuzzy because four different situations run under it. Separating them lets you measure and work on each one.
This is the ghost citation in the sense Kevin Indig gave it. Your own page appears as a source link under the answer. The brand name is missing from the answer text.
The user can click the source. In practice they rarely do. The brand stays a footnote to them.
The system used the content of the page. It reproduced that content without a sender. A typical answer sentence says that a recent analysis puts the figure at a certain value.
For measurement this case is the most convenient one. Your own domain stands in the source list and can be counted there.
Here the brand name stands in the answer text. A source reference for it is missing.
For perception this is the better case. The reader sees the name and attaches the statement to it.
For measurement it is the harder one. There is no address at which you could check where the mention came from. That is the gap between being named and being listed as a source.
This case arises in two ways. Either the name comes from model knowledge, or the system took it from a fetched page and did not link that page.
The answer itself does not tell the two routes apart. The test for it needs a second run with web search switched off.
In this case a link sits under the answer, and the linked page does not contain the claim attributed to it. The address works. The topic often even fits.
This gap between a cited source and a covered claim can be described as a verification gap. It has been measured, and the values appear further down in this article.
For a brand this case is the most awkward. Your own page stands as evidence under a statement that never appeared there.
Any reader who follows the link can notice it. Checking it takes opening the original.
Here the reference points at a work that does not exist. This is the research sense of the term.
Author, title, year and location look correctly formatted. The work itself does not exist. In Google Scholar a search for the title returns no hit.
Ghost references arise when a model produces a reference that never existed. It takes the format from its training data and invents the content to go with it.
This fourth case affects brands only indirectly. It does show how reliably a model handles references when nobody checks.
| Case | What the answer contains | What is missing | Field |
|---|---|---|---|
| Ghost citation | a source link to your own page | the brand name in the answer text | AI visibility |
| Mention without citation | the brand name in the answer text | a source reference for it | AI visibility |
| Verification gap | a link and a claim | cover for the claim in the source | AI answers and deep research |
| Ghost reference | a correctly formatted reference | the cited work itself | research |
Two studies from 2026 address the first case. They arrive at different figures, and the reason sits in the denominator.
The study records 3,981 brand appearances across 115 prompts. According to the authors the prompts were "run across 14 countries". Four AI engines were queried, namely ChatGPT, Google AI Overviews, Gemini and Google AI Mode.
The data comes from the Semrush AI Visibility Toolkit. The post appeared on 9 June 2026, written by Margarita Loktionova together with Kevin Indig and Growth Memo.
The dataset is small. 115 prompts are a manageable question list, and no collection period is given.
The prompts themselves are unpublished as well. That leaves the analysis impossible to recompute and impossible to compare with your own.
This is not specific to this one study. No AI visibility vendor we have checked publishes its question list.
Every brand appearance falls into exactly one of three states. The three shares add up to one hundred per cent.
Two totals follow from that. 74.9 per cent of all appearances contain a citation. 38.3 per cent contain a mention.
The headline of the Semrush post speaks of 62 per cent of AI citations. The three reported shares refer to appearances, because they add up to one hundred per cent.
Computed on the citations alone the value comes out differently. 61.7 divided by 74.9 gives around 82 per cent.
Both numbers describe the same dataset. Anyone putting them side by side should write the reference down with them. A version that does not hold reads "62 per cent of AI citations do not name the brand". The version that holds reads "61.7 per cent of all brand appearances are citations without a brand mention".
In roughly one case in eight both events coincide. Only there does the reader see the name and find the address behind it at the same time.
Those 13.2 per cent are the only class in which a mention is visible and traceable at once. Every other appearance delivers one of the two.
For practice that means a monitoring figure without this split says little. A vendor reporting 100 appearances can mean 100 footnotes or 100 named mentions.
On 29 July 2026 Search Engine Land published a second analysis. The data comes from Writesonic and covers a window of 30 days.
The scope is around 16 million brand appearances across seven AI engines. The result there reads "Around 40% of AI citations didn't name the source brand in the generated answer."
The authors are Nikki Lam and Samanyou Garg. The post discloses that Garg is founder and CEO of Writesonic.
Both studies measure the same event and arrive at 61.7 against around 40 per cent. Three differences explain the distance.
Neither figure is wrong because of that. They are two measurements taken on two populations. Comparing the percentages without that note misleads.
A fourth difference cannot be resolved. The Semrush analysis names no collection period, and the Writesonic data covers 30 days. Whether both captured the same phase of these systems is therefore open.
A version that does not hold reads "between 40 and 62 per cent of AI citations fail to name the brand". The version that holds reads "two studies from 2026 measure the share against two different populations, at 61.7 per cent of brand appearances and around 40 per cent of citations".
The largest finding in both studies is not the average. It is the spread between the systems.
In the Semrush analysis the citation rate for ChatGPT is 87 per cent. The mention rate is 20.7 per cent.
The system therefore links a source almost every time. It puts the brand name into the text in roughly one answer in five.
For a brand that means a high number of footnotes at a low presence of the name. Anyone counting citations alone reads ChatGPT as the friendliest channel.
Anyone counting the mention rate instead arrives at the opposite verdict. Both readings come from the same dataset.
The practical consequence concerns prioritisation. A team that runs ChatGPT as its most important channel should know that it appears there mostly as a source and rarely by name.
With Gemini the picture inverts. The mention rate is 83.7 per cent, the citation rate 21.4 per cent.
Gemini puts brand names into the answer text around four times as often as ChatGPT. It supplies a source link in roughly one case in five.
For a brand that is the more visible channel and the less traceable one at the same time. The name stands in the text, and the address behind it is missing.
Anyone adding Gemini to a monitoring setup therefore measures mostly names. Anyone running the citation rate as the lead metric reads Gemini as a weak channel, although most names fall there.
The two systems sit at opposite ends of the same scale. Between 20.7 and 83.7 per cent mention rate lies a factor of four.
Google AI Overviews sits between the two poles and leans towards the citation, as the authors describe it. Google AI Mode names brands roughly twice as often as ChatGPT in the same analysis.
For the two Google surfaces the post reports no individual percentages. That is a blank in the source and not an omission on our side.
A value that was never published cannot go into a table. We leave the row out rather than estimate it.
Writesonic reports for seven systems how often a citation stays without a brand name. Perplexity sits at the top of that list, Microsoft Copilot at the bottom.
| AI engine | Share of citations without a brand name |
|---|---|
| Perplexity | 52 % |
| Google AI Mode | 49 % |
| Google AI Overviews | 41 % |
| ChatGPT | 37 % |
| Gemini | 25 % |
| Grok | 22 % |
| Microsoft Copilot | 19 % |
The ranking backs the first study at one point. Gemini comes out clearly better than the average in both datasets.
At another point the two series diverge. ChatGPT reaches 37 per cent at Writesonic, while the Semrush values point at a far higher share. The different denominator explains part of that, and neither publication explains the rest.
The engine is not the only factor at work. The Semrush analysis reports three more that move the share of mentions noticeably.
The analysis puts short questions at a multiple of the mentions long questions produce. That is the largest single effect in the dataset.
A short question leaves the system little context. It answers with names, because a name is the shortest form a recommendation can take.
A long question supplies context along with it. The system answers by explaining and props the explanation up with links.
For informational questions the citation rate is 89.3 per cent and the mention rate 18 per cent. For comparison questions the mention rate rises to 43.3 per cent.
For how-to questions it sits at 42.8 per cent. For commercial questions the mention rate is 35.6 per cent and the citation rate 84.4 per cent.
| Question type | Citation rate | Mention rate |
|---|---|---|
| informational | 89.3 % | 18 % |
| comparison | not reported | 43.3 % |
| how-to | not reported | 42.8 % |
| commercial | 84.4 % | 35.6 % |
One practical consequence follows for every prompt list. Anyone querying informational questions only measures a band with a systematically low mention rate.
The mention rate varies considerably across the 14 countries in the study. India and Sweden sit at around 50 per cent.
Canada sits at 44 per cent and the United Kingdom at 41 per cent. Italy, Brazil and the Netherlands sit between 18 and 22 per cent.
The distance between the top and the bottom is therefore more than double. A target value from one market does not carry over to another.
| Market | Mention rate |
|---|---|
| India | around 50 % |
| Sweden | around 50 % |
| Canada | 44 % |
| United Kingdom | 41 % |
| Italy | 18 to 22 % |
| Brazil | 18 to 22 % |
| Netherlands | 18 to 22 % |
The analysis offers no explanation for that spread. Differences in language, in the volume of available sources and in the make-up of the prompts per market are all conceivable.
A plain rule follows for your own measurement. The market belongs in the analysis as its own field, and values from different markets are not averaged.
Germany does not appear in the published country list. Austria and Switzerland do not appear there either.
For the German speaking market there is therefore no ghost citation figure from a study with a published method. That is the position on 26 August 2026.
The same holds for every market outside the published list. Anyone who needs a figure for such a market measures it themselves. What that setup looks like is covered in the article on prompt tracking.
The third case is the least conspicuous of the four. An answer names a source, the link works, and the linked page says something else.
A paper titled "Cited but Not Verified" measured exactly that. It appeared on 7 May 2026 and examines source attribution in reports from deep research agents.
The authors describe the starting point like this. "Large language models (LLMs) power deep research agents that synthesize information from hundreds of web sources into cited reports, yet these citations cannot be reliably verified."
14 models from closed and open development were tested. Every single citation was assessed against the fetched content of the linked page.
The setup therefore differs from the usual fact check. A usual check assesses a statement on its own. This study fetches the linked page and holds the statement against it.
That measures exactly the case at issue here. A citation stands there, and the question is whether it carries.
The study rates every citation on three dimensions. The results per dimension fall far apart.
| Dimension | What is checked | Result for the strongest models |
|---|---|---|
| Link Works | whether the address is reachable | over 94 % |
| Relevant Content | whether the source fits the topic | over 80 % |
| Fact Check | whether the source covers the claim | 39 to 77 % |
The first two values are high. The third one is not.
The pattern is notable because it slips past exactly the check a reader usually performs. They click the link, see a page on the right topic and read no further.
The paper also examined what happens to accuracy when an agent researches more deeply. Tool calls were raised from 2 to 150.
Fact check accuracy falls by around 42 per cent across two frontier models in the process. The authors sum that up by stating that more retrieval does not produce more accurate citations.
A second side finding belongs with it. Fewer than half of the open source models managed to produce a report with citations at all in a one-shot setting. The original reads that "fewer than half of open-source models successfully generate cited reports in a one-shot setting".
Reports from deep research agents were measured. Those are long texts assembled over several research stages.
A normal answer in ChatGPT, Perplexity or an AI Overview is something else. For those short answers no comparable analysis with a published method exists.
The order of magnitude therefore does not transfer one to one. The direction of the finding stays usable. A cited source is not yet a checked source.
The second meaning belongs in a different context. It refers to a reference with no existing work behind it.
Anyone using the term in AI visibility should know the second meaning. It turns up in every search on the topic.
A paper called GhostCite measured the scale of the problem. Its subtitle reads "A Large-Scale Analysis of Citation Validity in the Age of Large Language Models". It was submitted on 6 February 2026 and revised on 14 May 2026. The authors work at Nankai University and Tsinghua University.
In the first part the authors had 13 large language models solve a citation generation task across 40 research domains. All 13 models invented sources.
The original reads "First, we benchmark 13 LLMs on citation generation task in various research domains, finding that all models hallucinate citations at rate from 14.23% to 94.93%."
The hallucination rate therefore runs from 14.23 to 94.93 per cent. The span between the best and the worst model is close to sevenfold.
An average across all models would carry no meaning here. The choice of model determines the result more strongly than any other factor in the setup.
In the second part the authors checked 2.2 million citations from 56,381 papers. The papers come from conferences in artificial intelligence, machine learning and IT security, NeurIPS among them, covering the years 2020 to 2025.
1.07 per cent of the papers contain at least one invalid citation. For 2025 the paper reports a rise of 80.9 per cent.
The absolute share is small. The movement inside a single year is not.
In the third part the authors surveyed 97 researchers. 87.2 per cent said they use AI tools in their work.
76.7 per cent of reviewers say they do not check the references of submitted papers thoroughly. 74.5 per cent consider peer review unsuited to finding citation errors. 41.5 per cent take BibTeX entries over without checking them.
The conclusion of the paper reads "Based on these findings, we argue that ghost citations represent a systemic threat to academic integrity."
These self reports are not a direct measurement. They describe what respondents say about their own behaviour, which puts them on a different evidence level from the 2.2 million checked citations.
They are worth reading all the same, because they explain the mechanism. An invented reference stays in circulation for as long as nobody opens it.
An invented reference looks correct. Author, title, journal and year sit in their places, and the format is right.
The check therefore runs through existence. In Google Scholar a search for the exact title either returns a hit or it does not. Every hit there also carries a "Cited by" line with the number of citations.
The same move works on a citation in an AI answer. Open the linked page and look for the quoted passage in the original.
Both meanings describe a hole in a reference. The location of the hole differs between them.
With the ghost citation in AI visibility the source is real and the name is missing. With the ghost reference in research the source is missing.
For a brand the first case is the one that matters in practice. It affects every answer in which its own page serves as evidence.
Up to here this has been about figures. From here it is about what they mean for your own visibility.
For many users an AI answer replaces the first look at a results list. What stands in that paragraph shapes the picture they take away.
If no name stands there, no picture forms. The statement works, and the brand behind it does not.
After the answer the user remembers an argument. They remember no vendor.
That hits the phase in which a prospect does not yet know the vendors. They ask an unbranded question, get an answer and take a list of names from it. A brand missing from that list barely features later on.
The effect can be reproduced in your own category. Put the typical entry question of your audience to a system and count which names appear in the text.
The second case turns the problem around. The name stands in the text, and an address for it is missing.
For perception that is the better outcome. For the analysis the point of attachment is gone.
Without a source link there is no way to determine which page on the web carried the mention. That removes the feedback about which placement worked.
A citation supplies an address and with it a starting point. A mention without a citation supplies none.
In practice we regularly see two reports side by side that draw different pictures of the same brand. One counts footnotes, the other counts names.
In our experience it pays to keep both figures separate from the start. They respond to different work, and they develop at different speeds.
Anyone introducing the split later loses comparability in their own time series. Old runs cannot be pulled apart retroactively, because answer texts are rarely archived.
The mention as a metric in its own right is covered in the article on AI mentions. It also lists which fields are worth recording beside the bare name.
Show me a single visibility percentage and I cannot tell you whether the brand was cited or named. I want both figures in the same table, because a campaign usually moves one of them before the other.
The measurement takes little effort once the fields are fixed. It becomes expensive only when somebody has to pull two values apart afterwards that were recorded together.
The following details are held for every answer a system gives. This is the minimum set needed to separate citation and mention later.
The last field costs the most effort. It pays as a sample, not for every answer.
The first six fields can be filled automatically. The seventh asks somebody to open the linked page and look for the passage.
A sample of ten answers a month is enough for an order of magnitude. If a claim turns up attributed to your own page that does not stand there, that belongs in the analysis.
The citation rate is the share of answers in which your own domain appears as a source. The mention rate is the share of answers in which the brand name stands in the text.
Both values belong in the same table and in separate columns. A third column records in how many answers the two occur together.
A version that does not hold reads "visibility in AI answers 34 per cent". The version that holds reads "citation rate 28 per cent, mention rate 12 per cent, both together 6 per cent, measured across 40 prompts in four runs". Those three percentages are an example, not a measured result.
A single run carries no statement. The same question draws different answers even in the same moment.
The variance in these systems is wide enough to cover small changes completely. How many runs are needed and what variance is normal is covered in the article on prompt tracking.
The ghost citation rate behaves like any other metric from these systems. It carries as a time series with a constant setup.
Two vendors can report different percentages for the same brand on the same day. The reason sits in the question list and in the selection of systems.
No vendor we have checked publishes its prompt set. That leaves no figure comparable between two vendors.
The full treatment of that non-comparability sits in the article on share of model.
A measurement says how often something occurs. It does not say why.
Whether a mention comes from model knowledge or from a web search at runtime cannot be read off the answer. The test for it sits in the article on LLM visibility.
The duration of every measure hangs on that origin. A gap in retrieval can be worked on in days. A gap in model knowledge cannot.
Nobody outside a lab has direct access to a model's training data. It changes with the next training run, and those runs come months apart.
The citation rate hangs on your own website. The mention rate hangs on what other pages write about the brand. Those are two different jobs.
A page from which single statements can be lifted cleanly gets pulled as a source more often. That takes a clear definition, dated key facts and an unambiguous assignment of the brand to its category.
Semrush draws a section from its own analysis under the heading "How to close the citation gap". The recommendation there is to place brand and product mentions more clearly in your own content when a page gets cited without the brand being named.
This work lands more reliably on the citation rate than on the mention rate. A system that leaves the name out of the answer text does so by its own writing logic.
The mention rate responds to what stands about the brand outside your own domain. A system assembling an answer from several sources finds more evidence for a brand name the more independent pages carry that name in the same context.
Exactly that evidence is built by brand mentions and digital PR on suitable publisher pages. A trade article naming the brand in the body text hands the system a name in the text and no bare address.
A brand appearing in a trade publication, in a study and in an interview with the same focus supplies three independent pieces of evidence for the same assignment. That is the kind of basis a system takes a name into an answer text from.
Mentions on the web correlate with mentions in AI answers. The relationship is not causally proven. There is so far no controlled experiment that attributes one additional placed mention to one additional AI mention. Anyone promising a guarantee here does not have the evidence for it.
An editorial placement produces two things at once. It puts the brand name into a text, and it puts a fetchable page on the web that ties that name to a statement.
For the retrieval route the fetchable page is what counts. For the answer text what counts is that the name stands there written out.
A weak placement brief asks for a "mention of the client in the author profile below the article". The stronger brief asks that the brand name stand in the body text of the article, inside a sentence that contains a checkable detail about the company.
An example of such a sentence from a trade article reads "the provider GetMentioned lists publishers in several languages according to its own account and shows the price per placement before booking". The name stands in it, and the sentence carries a detail somebody can check.
A sentence like that is easier for a system to take over than a list of vendor names with nothing added. It ties the name to a property.
The answer itself cannot be booked. These systems work probabilistically, and the same question can return two different answers in two runs.
Whether a system puts a name into the answer text is decided per answer. No vendor can commit to it.
What can be procured is the basis for it. That means pages on which the name stands in a fitting context.
GetMentioned is a link building agency with its own placement marketplace. Full service in link building, not full-service SEO.
Our work starts at the second job, so at what stands about a brand outside its own domain. Tracking individual answers in answer engines is not part of our product.
Not every placement carries a brand name into a text. We check these features before booking.
The GetMentioned Score brings price, SEO data and quality signals together per publisher. It does not replace the editorial check on topical fit.
In the marketplace publishers stand side by side with SEO data, topic assignment and price. The price you see in the marketplace is the price you pay.
MentionIQ analyses your website, your competitors and your important keywords and proposes suitable link sources from that. It narrows the selection down to publishers already visible in your own topic area.
Teams that want to hand the work over can give it to us as a managed mandate. Then we take on analysis, publisher selection, text, coordination, publication and monitoring for you.
The text is the place where the mention gets decided. We coordinate it so that the brand name stands in the body text and is tied to a statement there.
In our experience that is the point at which many campaigns give away their effect. The link sits cleanly, and the name appears nowhere in the text.
Work on the client's website is outside the trade. Site structure, technical SEO and load time stay with the client or with their SEO support.
We write and deliver SEO texts, including for the client's own pages. The client publishes them there.
Measuring your own AI visibility is outside it as well. There are dedicated vendors for that, for example Peec.ai or Rankscale.
Four sentences circulate about the ghost citation problem that the available data does not support. The defensible version follows each one.
The data shows the opposite of a fixed coupling. ChatGPT delivers 87 per cent citations and 20.7 per cent mentions.
A rise in citations can leave the mention rate unchanged. The defensible version reads "citations and mentions develop independently of each other in the systems tested, and both belong in separate columns".
25.1 per cent of all brand appearances in the Semrush analysis are mentions without a citation. For the reader that is precisely the case in which the name becomes visible.
The defensible version reads "a mention without a link reaches the reader, and it gives the analysis no address at which its origin could be checked".
Two studies from the same year arrive at 61.7 and at around 40 per cent. The per engine values run from 19 to 52 per cent.
The defensible version reads "the ghost citation rate hangs on the engine, on the question type, on the market and on the chosen denominator, and a single percentage without those four details cannot be interpreted".
The study on deep research agents reports link reachability above 94 per cent and fact check accuracy from 39 to 77 per cent. A working link therefore says little about cover for the claim.
The defensible version reads "a source link documents that a page is reachable, and whether the page carries the claim attributed to it shows only in a sample checked against the original".
The ghost citation problem is a measurement problem at its core. A citation and a mention are two events, and in the systems tested they mostly turn up one at a time.
Anyone folding both into one figure loses the information every measure depends on. A report with a separate citation rate and mention rate says more in one line than a visibility score in per cent.
The citation rate responds to your own website. The mention rate responds to what other pages write about the brand. In practice we see the second gap more often, and it is the more awkward one, because it cannot be closed inside your own content management system.
A ghost citation is a source reference in an AI answer that comes without the brand name in the answer text. The page appears as a source link, and the reader never reads the name behind it.
Semrush credits Kevin Indig with coining it, in the words "Kevin Indig coined the term ghost citation to describe this gap". The joint analysis appeared on 20 April 2026 at Growth Memo and on 9 June 2026 at Semrush.
In the analysis by Semrush and Kevin Indig, 61.7 per cent of all brand appearances are citations without a brand mention. A second study by Writesonic across around 16 million brand appearances arrives at around 40 per cent of citations.
A citation is a clickable source link in or under the answer. A mention is the brand name in the answer text. Only 13.2 per cent of the appearances tested carry both.
Of the four systems tested, Gemini names brands most often, at a mention rate of 83.7 per cent against a citation rate of 21.4 per cent. ChatGPT runs the other way round at 87 per cent citations and 20.7 per cent mentions.
There the term refers to a reference with no existing work behind it, more often called a ghost reference. A study from February 2026 measures hallucination rates of 14.23 to 94.93 per cent for that across 13 large language models.
The published country list covers 14 markets, among them the United Kingdom at 41 per cent and Canada at 44 per cent. Germany, Austria and Switzerland do not appear in it. For a market outside the list the figure has to come from your own prompt list.
About the author
Managing Director, GetMentioned
David has been building link acquisition and digital PR processes since 2016, first as an agency under SEO Galaxy, today as a platform with GetMentioned. He has scaled his own projects from zero to seven-figure monthly traffic and delivered thousands of campaigns for clients. Here he writes about what works in practice, and about what only costs budget.
Compare publishers, SEO data and prices in one place and book the placements that fit.
Share of Model tracks a brand's share of all brand mentions in AI answers. Why the formula is uncontested but the denominator is not, and why two vendors can report different values for the same brand on the same day.
August 23, 2026 · 10 min read
A brand surfaces either from model weights or from a web search at runtime. A half-hour test separates the two, and that decides whether the work takes days or months.
August 16, 2026 · 12 min read
The same question returns different answers at the same moment. What a measurement setup that survives this looks like, and why numbers from two tools cannot be compared.
August 15, 2026 · 15 min read
