Ghost Citations: Why AI Can Cite Your Website Without Mentioning Your Company

That is different from a hallucinated citation. A hallucinated citation is fabricated, nonexistent or materially invalid. A ghost citation in the sense used here involves a real source attribution with an absent company or brand mention.

The term has gained traction in AI-search marketing during 2026. Semrush uses it for a citation where the source brand is absent, while Search Engine Land and Writesonic use the same definition in their recent analysis. (semrush.com)

The distinction matters because a source link and company visibility answer different questions.

Kojable's Citation Rank and Brand Visibility research examines that distinction directly. The study tracks whether a cited source's associated entity actually appears in the generated prose. Even among entities associated with the first cited-source position, most remained unnamed. The research is observational and finance-specific, so it does not establish that citation position caused an entity to appear or disappear.

That leads to a more useful measurement sequence:

1. Was the source visibly cited?
2. Was the associated company or entity named?
3. If named, how was it described or framed?
4. Was it compared with competitors?
5. Was it actually recommended?

Those outcomes should not be compressed into one “AI visibility” result.

A company can therefore be:

• cited and named;
• cited but unnamed;
• named without its website being cited;
• neither cited nor named.

Kojable's AI citations Reference Entry treats citation and company mention as separate evidence objects for exactly this reason. It also cautions against renaming a visible citation as hidden retrieval, grounding, influence or support.

That is evidence worth recording. It is not a complete explanation of why the answer occurred.

A visible citation does not automatically prove that the cited page:

• caused a claim to appear;
• was the most influential source;
• was first in an internal retrieval process;
• was trusted or endorsed by the AI system;
• was used in model training;
• was the reason a competitor was named;
• should be the first page you change.

Kojable's citation research makes the same distinction between source inclusion, entity presence and downstream recommendation. Its separate recommendation study found that citation position should not be interpreted as endorsement.

This is also why phrases such as “the AI used our content but hid our brand” should be treated cautiously. A citation may make the source relationship visible to the reader, but the citation itself is not internal model telemetry.

The practical rule is simpler:

Record what you observed first. Diagnose the explanation separately.

It may matter little in a neutral educational answer and substantially more in a vendor, comparison, evidence-ownership or recommendation question.

Consider the difference between these questions:

These are decision rules, not a validated numerical scoring system.

A company does not automatically deserve a mention because one of its pages was cited. In a purely educational answer, naming every source owner may add little value.

The gap becomes more important when the question itself requires company identification, comparison, differentiation, evidence ownership or recommendation.

That is the point where a ghost citation becomes part of a wider AI representation problem rather than simply a citation observation.

Use separate fields for the separate outcomes.

A fabricated or nonexistent reference belongs in a separate category. It is a citation-validity or hallucination problem, not simply a citation-without-company-mention state.

For each buyer-relevant observation, a useful record includes:

• exact question;
• AI platform and surface;
• date and run;
• whether the company was named;
• whether a company source was cited;
• exact cited URL;
• competitors named;
• whether the company description was accurate;
• whether the company was recommended where recommendation was relevant.

This keeps the measurement tied to the actual buyer question instead of reducing several outcomes to one score.

Recent studies all show that citations and brand mentions can diverge, but they report materially different rates because their samples, platforms, prompts and analytical units differ.

Semrush records whether each domain appearance was cited, mentioned or both. (semrush.com) Search Engine Land and Writesonic explicitly describe their result as a snapshot and state that the study does not establish the causes of ghost citations or their effects on clicks, recall, trust or conversions. (searchengineland.com) BuzzStream uses yet another design and counts the overlap between brand-domain citations and brand mentions. (buzzstream.com)

These studies should not be averaged together.

Their disagreement is useful because it demonstrates why a company's own measurement contract matters more than a generic market benchmark.

The better question is:

Across the buyer questions that matter to us, how often are we cited, named, represented accurately and recommended under clearly defined conditions?

Start with the buyer-facing gap, not with a theory about hidden model behaviour.

A useful diagnosis asks six questions.

1. Did the question actually call for a company mention?

Review the exact prompt and answer first.

If the question asks for a neutral definition, an absent company may be expected. If it asks for vendors, alternatives, providers or evidence ownership, the same absence may be much more important.

2. Does the pattern recur?

One answer is an observation.

Repeated comparable answers across relevant prompts or surfaces can establish a recurring pattern, but recurrence still does not prove why the pattern exists.

Keep the prompt, surface and measurement conditions comparable before calling something recurring.

3. Is the company clearly connected to the evidence on the cited page?

Inspect the cited page itself.

Ask:

• Is the company clearly identified?
• Is the relevant claim connected to the company or research owner?
• Is the information current?
• Does the page contain specific, verifiable evidence?
• Does the relationship between entity, claim and evidence require unnecessary inference?

Kojable's citation-rank research identifies clear entity identification and explicit entity-to-claim relationships as defensible information-quality practices. It does not establish them as guaranteed levers for causing a future company mention.

4. Is the evidence the buyer would need actually available?

A company may have a technically clear page while still lacking the evidence required for the buyer question.

For example:

• the page explains the category but not the company's differentiated position;
• the company claims enterprise suitability but offers little current enterprise proof;
• the website reflects current positioning while external descriptions still reflect an older offer;
• competitors provide clearer evidence for the decision being asked about.

That is an evidence problem, not simply a naming problem.

5. What do the other cited sources say?

AI answers can contain several source families at the same time.

A company-owned page, publication, review, directory and competitor source may all appear in the observed citation environment. Their presence tells you what information is worth inspecting. It does not reveal which one had the greatest hidden influence.

Review whether those pages contain:

• current or outdated positioning;
• relevant independent evidence;
• different category language;
• competitor-led framing;
• conflicting descriptions.

Do not call a source authoritative merely because it recurs.

6. Are competitors being surfaced where your company is not?

This is often the commercially important part of the diagnosis.

If your page is cited in a category answer while several competitors are named and your company is absent, record the pattern.

Then ask what evidence those competitors are associated with and whether your own relevant positioning or proof is missing, unclear or inconsistent.

Do not jump from “competitor named” to “competitor source caused our omission”.

That distinction is central to Answer Intelligence: separate the observed answer pattern from the evidence and plausible conditions worth investigating.

A cited URL is evidence. It is not an instruction telling you where the remediation must happen.

Start with the material problem:

• What should the buyer understand?
• What is missing, inaccurate or poorly evidenced?
• Which information surface actually owns that truth?
• What is the smallest justified intervention?

Sometimes the cited page will be the right place to act. For example, the company-to-claim relationship may genuinely be unclear or the page may contain outdated information.

In other cases, the justified action could involve:

• adding missing proof to another owned page;
• updating current company positioning;
• correcting an outdated third-party profile;
• creating evidence that does not yet exist;
• clarifying a comparison or use case;
• doing nothing because the company did not need to appear in that answer.

Kojable's AI Representation Remediation Guide owns the full intervention process. Its central principle is to remediate the diagnosed representation gap rather than treating citation presence itself as the problem.

For a ghost-citation problem, the target might be:

“On this buyer question, does the company now appear accurately in the answer while its relevant evidence is still represented appropriately?”

Track citation presence and company/entity presence separately. Where relevant, also track:

• description accuracy;
• category or audience framing;
• competitor co-mentions;
• recommendation status;
• cited-source changes.

Do not turn a single changed answer into a causal claim.

A before-and-after difference shows that the recorded answer changed. Stronger statements require stronger evidence and repeated comparable observations.

Kojable's AEO Change Verification method recommends frozen longitudinal prompts, predefined outcome measures and observation checkpoints rather than fixed “propagation deadlines”. It explicitly separates observed movement from causal attribution.