Ghost citations: why AI cited your content but forgot your brand

A source link in an AI answer is not the same as a brand recommendation. Semrush and Growth Memo measured 3,981 domain appearances across 115 prompts, 14 countries, and four AI search engines. The result: 61.7% of AI citations are ghost citations, meaning the domain appears as a source link but the brand name is absent from the answer text. ¹²⁶

That distinction changes the scoreboard. A GEO dashboard that reports only "cited" will overstate visibility. A brand can be cited as a footnote while the AI recommends a competitor by name. The content worked; the brand did not.

The three states of AI visibility

Every AI appearance should be classified into one of four states. The useful metric is not citation count alone; it is the split between citation, mention, and recommendation.

The platform split is not cosmetic. Seer’s earlier 541,213-response study found the same citation-vs-mention gap across 20 brands, and Everything-PR’s 680M+ citation index shows how heavily AI source concentration favors a small set of domains. ³⁵ Hidden Layer should not treat “cited” as a single bucket. It should split citation, mention, and recommendation before making any strategic claim.

Why the platforms behave differently

The ghost citation problem is platform-specific. Gemini names brands far more often than it links sources. ChatGPT does the reverse: it links sources but often omits the brand name from the answer body. That difference matters because a ChatGPT-first strategy and a Gemini-first strategy are not the same strategy.

CiteMetrix found only 11% domain overlap between ChatGPT and Perplexity top-cited domains, confirming that platform-specific source behavior is not a minor implementation detail. ⁴ A brand can be visible in one engine and absent in another without doing anything wrong. The measurement system has to know which engine it is measuring.

Citation selection is not citation absorption

There are two stages in the citation path. First, the AI system selects a source during retrieval. Second, the generated answer absorbs that source in a way that affects the final response. A ghost citation is usually a failed absorption problem: the page was selected, but the brand did not become part of the answer.

• Selection failure: the page is not retrieved at all. This is a crawlability, entity, authority, or freshness problem.
• Absorption failure: the page is retrieved and cited, but the brand is not named. This is a content-positioning and brand-memory problem.
• Recommendation failure: the page is cited, the brand is named, but the AI recommends another provider. This is a differentiation and proof problem.

Most GEO advice stops at selection. It says "get cited." Hidden Layer should measure all three stages. A source link without a brand name is not a win. It is evidence that your content is useful enough to be used and weak enough to be forgotten.

The ghost citation rate formula

Use a simple formula and keep it separate from citation share.

Ghost citation rate = ghost citations / total AI appearances
Cited + mentioned rate = appearances with brand name and source URL / total AI appearances
Mention rate = appearances naming the brand / total AI appearances
Recommendation rate = appearances naming the brand as an option or recommendation / total AI appearances

The denominator matters. If a prompt set has 100 brand appearances and 62 are ghost citations, ghost citation rate is 62%. If the same brand has 500 total query runs and only 100 appearances, appearance rate is 20%. Both numbers are useful. One measures quality of appearances; the other measures reach.

How to reduce ghost citations

1. Name the brand in the answer block. The first 40–60 words of a citable section should state who wrote it and what it proves.
2. Turn generic claims into attributable claims. "Strong PIM governance improves AI product recall" is weaker than "Hidden Layer audits show PIM-first catalogs score higher on product-page completeness."
3. Add primary data. Original audit data, benchmark tables, and methodology sections give AI systems a reason to cite the brand, not just the URL.
4. Use comparison pages carefully. "Brand A vs Brand B" pages should state the differentiator in the prose, not only in tables.
5. Track mention rate separately from citation rate. If citation rate rises while mention rate falls, the content is becoming a footnote factory.

The practical fix is not keyword stuffing the brand name into every paragraph. That is how pages become unreadable and spam-like. The fix is to make the brand part of the evidence: named methodology, named data, named author, named product, named category claim.

What Hidden Layer should measure

The Hidden Layer audit implication

The current audit already separates infrastructure from presence. Ghost citations belong in the presence layer. A domain can have perfect bot access, clean `llms.txt`, and valid schema while still losing brand memory to Reddit, Wikipedia, review sites, or a stronger competitor. That is not a technical failure. It is a citation-quality failure.

The next Hidden Layer measurement layer should add a citation probe to the audit result: run the domain against a query set, classify each appearance as cited, mentioned, ghost cited, or recommended, then compare the result to competitors. The output is not a grade for "SEO hygiene." It is a map of where the brand is visible, where it is invisible, and where it is being used without being remembered.

What to do next

Pick 25 prompts that a buyer would actually ask. Run them across ChatGPT, Perplexity, Google AI Overviews, and Gemini. Classify every appearance as cited, mentioned, ghost cited, or recommended. Then sort the results by prompt type and competitor.

If your citation rate is high and your mention rate is low, rewrite the cited pages so the brand becomes part of the evidence. If your mention rate is high and your citation rate is low, improve entity links, author schema, and source-quality signals. If both are low, start upstream: entity presence, authority, and retrieval access. The fix depends on the failure mode.

Run a Hidden Layer audit to see the infrastructure side of the same problem: bot access, entity signals, structured data, and product-page completeness. Then add the citation probe to measure the brand-memory side. Together, they show whether AI can reach you, understand you, cite you, and remember you.