Why AI Cites Your Content and Recommends Your Competitor

The Concept

The Citation Decision

When an AI search system retrieves your content, you have not earned a citation. You have earned a seat in the waiting room.

Retrieval is the first filter. Citation is the last. Between them sits a multi-stage evaluation pipeline that discards the majority of what it retrieves. AirOps analyzed 548,534 retrieved pages across 15,000 prompts and found that only 15% appeared in the final response. Ahrefs' study of 1.4 million prompts reported a higher aggregate figure, roughly 50%, but that number is inflated by compositional artifacts in the data: Reddit API feeds account for 67.8% of the non-cited pool and are cited at only 1.93%, dragging the denominator.

The directional finding is consistent across both studies: most content that makes it into the retrieval set is never cited. The model retrieves broadly and cites narrowly. Understanding what happens between those two steps is the subject of this issue.

ELI5: The Staff Meeting Analogy

Imagine a manager preparing a recommendation for the CEO. She asks her team to pull together everything relevant to the decision. Six people bring research. She reads all of it. But when she writes the memo, she cites two sources by name, paraphrases a third without credit, and ignores the rest entirely.

The three people who got ignored did real work. Their research was relevant enough to be pulled into the room. But it did not survive the manager's editorial judgment about what was specific enough, trustworthy enough, and clearly stated enough to attach her name to.

AI citation works the same way. The retrieval step is the team pulling research. The citation decision is the manager writing the memo. Your content can be in the room and still be invisible in the output.

The part that should concern most brands: one of the sources the manager paraphrased without credit was the most useful document in the stack. The manager already knew the main point from her own experience. She used that source to fill in a detail but did not feel the need to cite it because the conclusion was already hers. That is the ghost citation problem, and it is measurable.

Practitioner Level

The citation pipeline has four distinct failure points. Content can be eliminated at any stage, and each stage requires a different response.

Stage 1: The retrieval trigger.

Before any citation can happen, the model must decide to search. Not every query triggers retrieval. Semrush's analysis of 80 million clickstream records found that approximately 54% of ChatGPT queries are handled from parametric knowledge without triggering web search at all. Queries that include dates, price constraints, comparison structures, or requests for current data trigger retrieval at much higher rates. If the model decides it can answer from memory, no citation is possible regardless of how well your content is structured.

This is the stage where parametric presence matters most. If the model has a confident answer from training data, it may never look for yours.

Stage 2: Fan-out and candidate retrieval.

When retrieval is triggered, the model decomposes the query into sub-queries and runs them against its search index. AirOps found that ChatGPT generated two or more fan-out queries on 89.6% of searches, expanding 15,000 original prompts into 43,233 queries. Your content must match these specific sub-queries to enter the candidate set. A page optimized for the parent query may not match any of the sub-queries the model actually runs.

This is why passage-level retrieval readiness matters more than page-level optimization. The relevant update since Issue 01: Ahrefs measured cosine similarity between page titles and both the original prompt and the fan-out queries. Fan-out-query-to-cited-title similarity averaged 0.656 for cited pages, versus 0.602 for prompt-to-cited-title. Non-cited pages averaged 0.484 against the prompt. Content that aligns with the sub-questions the model actually asks cites better than content optimized for the parent query.

Stage 3: The gatekeeping layer.

This is the stage most practitioners skip entirely because it is the least visible. Between "your page appears in the search results" and "the model reads your page" sits an evaluation step where the model assesses title, snippet, and URL against its fan-out queries to decide which candidates are worth opening.

The model does not read every page it retrieves. It triages. ZipTie.dev's reverse-engineered analysis of Perplexity's pipeline describes a multi-layer reranking stage using cross-encoder models that evaluate semantic alignment between the query and the candidate content. Perplexity's own documentation confirms cross-encoder reranking as a documented pipeline stage. The surviving candidates are embedded into the prompt assembly before the language model generates the response, meaning citations are structurally bound to the generation process rather than appended afterward.

For ChatGPT, the reranking architecture is not publicly documented but is structurally confirmed: the OpenAI API distinguishes between "sources" (all URLs consulted) and "url_citation" annotations (only the most relevant references), confirming that a selection step exists between retrieval and citation.

The practical implication: your page title and meta description are not just click-through optimization for human searchers. They are the first signal the model uses to decide whether your content is worth reading at all. A title that clearly states what question the page answers, aligned with the vocabulary of likely fan-out queries, is mechanically advantageous at this stage.

Stage 4: Passage-level extraction and citation attachment.

The model has now opened the page. It evaluates the content at the passage level, not the page level. It is looking for a specific, extractable passage that answers the sub-query it retrieved for. If your page buries the answer under introductory context, or if the answering passage mixes in unrelated information, the model may absorb the general insight without citing the source. This is where semantic density at the section level directly affects whether your content earns the citation or just informs the answer anonymously.

The GEO-SFE paper (arXiv:2603.29979, March 2026) provides the most specific structural data on what survives this stage. Chunks exceeding 300 words showed 31% attention degradation. Structured formats like lists and tables showed 43% higher extraction accuracy than equivalent prose. Sentence-initial positions received 2x attention magnitude compared to mid-sentence positions. The data supports what the mechanism predicts: passages that lead with the claim and contain it within a semantically dense, structurally bounded chunk are more likely to be both extracted and cited.

The GEO-16 framework study (arXiv:2509.10762) audited 1,100 URLs across three engines and found that pages with strong metadata, semantic HTML, and structured data had the highest association with citation likelihood. Pages achieving a GEO score above 0.70 and hitting 12 or more quality pillars achieved a 78% cross-engine citation rate. This is the mechanical basis behind citation-ready content as a practice: engineering passages to survive each stage of this pipeline, not just to read well.

The Technical Layer

Ghost Citations: When Retrieval Succeeds and Visibility Fails

The most important finding in AI citation research in 2026 is not about what gets cited. It is about what gets cited without producing any brand visibility.

Seer Interactive analyzed 541,213 LLM responses across 20 brands and 6 AI platforms and identified a pattern they call the ghost citation. A ghost citation occurs when a brand's URL is cited as a source but the brand itself is never mentioned by name in the response text. In the most damaging variant, the competitive ghost citation, the brand's content is cited while a competitor is explicitly named and recommended in the same response.

The numbers are stark. When a brand is mentioned in a response, its citation rate is 53.1%. When the brand is not mentioned, the citation rate drops to 10.6%. That is a 5x differential, and it runs in the wrong direction for anyone who assumes citations drive mentions.

Growth Memo's independent analysis, using Semrush AI Toolkit data across 3,981 domain appearances, corroborates the pattern: 61.7% of all citations were classified as ghost citations. Only 13.2% of domain appearances converted into both a citation and a brand mention.

Why ghost citations happen: the post-hoc citation hypothesis.

Seer Interactive's leading hypothesis, backed by six independent behavioral tests across 362,188 LLM responses, is that the causal direction is backwards from what most practitioners assume. The model does not read your content, find it persuasive, and then recommend your brand. The model recommends brands it already knows from parametric memory first, then retrieves content to support the recommendation after the fact. The citations are post-hoc justification for a decision the model already made.

Seer is appropriately transparent about the limits of this finding: they do not have access to token generation logs and cannot observe the internal sequence of operations directly. This is behavioral evidence, not architectural proof. But the behavioral signal is consistent across their dataset. Passionfruit Labs documented a related phenomenon: citation volatility so high that a citation appearing in one response has less than a 1 in 100 chance of appearing identically across multiple prompt runs. This is consistent with a system where the citation is a post-hoc attachment to a parametrically determined answer rather than a stable input to the answer itself.

The implication is structural. If you want your brand cited, you optimize content for retrieval: structure, semantic density, passage-level answerability. If you want your brand mentioned and recommended, you need parametric presence: entity signals in the Knowledge Graph, Wikipedia coverage, consistent naming across authoritative third-party sources, brand-as-subject positioning in your own content. These are two different optimization targets with two different timelines. Seer found that content changes propagated to retrieval systems within days. Brand mention changes took six to twelve weeks.

The ref_type dimension: where citations actually come from.

Ahrefs' April 2026 study of 1.4 million prompts revealed that ChatGPT categorizes its retrieval sources by channel using an internal ref_type field, and the citation rates vary dramatically by channel.

The general search index accounts for 88.46% of all citations. News feeds account for 12.01%. Reddit is retrieved at massive volume but cited at only 1.93%. YouTube is cited at 0.51%. Academic sources at 0.40%.

The Reddit number is instructive. ChatGPT retrieves Reddit extensively to gauge consensus, calibrate sentiment, and build context for its answers. But it almost never cites Reddit as a source. It learns from the crowd, then cites an institution. Any analysis comparing "cited vs. non-cited" URLs without controlling for ref_type is measuring compositional artifacts, not real citation signals.

Platform Differences

How the citation decision varies across platforms

The ghost citation data from Growth Memo exposes a platform divergence that should change how practitioners allocate optimization effort.

Dimension	ChatGPT	Perplexity	Google AI Overviews	Gemini
Citation rate	87.0% of appearances include a citation link	Highest citation density; 5+ sources per response typical	84.9% citation rate; avg 15.22 sources post-Gemini 3	21.4% citation rate
Brand mention rate	20.7%	Not separately measured in Growth Memo dataset	61.0%	83.7%
Dominant behavior	Cites frequently, mentions rarely. Operates like an academic paper with footnotes.	Retrieval-first; most correctable platform because parametric layer is least dominant	Closest balance between mentions and citations	Mentions frequently, cites rarely. Operates like a conversationalist drawing on brand knowledge.
Ghost citation risk	High. Your content provides the evidence; your brand is absent from the recommendation.	Moderate. Retrieval-first architecture means retrieved content has more influence on the answer.	Moderate. Knowledge Graph presence feeds both citation and mention.	Low for citation ghosts, high for mention-without-citation. Different problem: brand gets named but no link drives traffic.
Optimization priority	Passage-level retrieval readiness + parametric entity signals	Semantic density + freshness + structural extractability	Entity SEO + Knowledge Graph presence + traditional authority signals	Parametric presence is primary; citation optimization secondary

Dimension

Citation rate

ChatGPT

87.0% of appearances include a citation link

Perplexity

Highest citation density; 5+ sources per response typical

Google AI Overviews

84.9% citation rate; avg 15.22 sources post-Gemini 3

Gemini

21.4% citation rate

Dimension

Brand mention rate

ChatGPT

20.7%

Perplexity

Not separately measured in Growth Memo dataset

Google AI Overviews

61.0%

Gemini

83.7%

Dimension

Dominant behavior

ChatGPT

Cites frequently, mentions rarely. Operates like an academic paper with footnotes.

Perplexity

Retrieval-first; most correctable platform because parametric layer is least dominant

Google AI Overviews

Closest balance between mentions and citations

Gemini

Mentions frequently, cites rarely. Operates like a conversationalist drawing on brand knowledge.

Dimension

Ghost citation risk

ChatGPT

High. Your content provides the evidence; your brand is absent from the recommendation.

Perplexity

Moderate. Retrieval-first architecture means retrieved content has more influence on the answer.

Google AI Overviews

Moderate. Knowledge Graph presence feeds both citation and mention.

Gemini

Low for citation ghosts, high for mention-without-citation. Different problem: brand gets named but no link drives traffic.

Dimension

Optimization priority

ChatGPT

Passage-level retrieval readiness + parametric entity signals

Perplexity

Semantic density + freshness + structural extractability

Google AI Overviews

Entity SEO + Knowledge Graph presence + traditional authority signals

Gemini

Parametric presence is primary; citation optimization secondary

The practical read: a brand that is well-cited on ChatGPT and invisible on Gemini may have strong retrieval optimization and weak entity signals. A brand that is frequently mentioned on Gemini but rarely cited anywhere has strong parametric presence and weak content structure. The diagnosis determines the prescription.

What Changed Recently

February to April 2026 developments worth knowing

Seer Interactive's ghost citation research (March 2026) introduced the most useful new metric in AI visibility: the competitive ghost citation rate, measuring how often your content is cited in responses that recommend a competitor. This is directly quantifiable revenue exposure. In their dataset, it was measurable in every sector analyzed.

The GEO-SFE paper (March 2026) provided the first peer-reviewed structural optimization benchmarks specifically for AI citation. Prior to this, structural recommendations for AI citation were practitioner-derived observations. The GEO-SFE data gives specific targets: 150-300 word paragraph length, 3-5 levels of heading depth, 25-35% structured elements, answer-first positioning. These are not aspirational guidelines. They are the structural properties that empirically predict citation survival.

Fischman's schema markup study (February 2026, preprint) found that generic schema (Article, Organization, BreadcrumbList) provides zero measurable AI citation advantage. Attribute-rich schema (Product and Review with populated pricing, ratings, and specifications) outperformed generic schema by approximately 20 percentage points in citation rate, but the advantage was concentrated in lower-authority domains. For domains with high existing authority, schema type made little difference. The implication: schema is a tiebreaker for the middle of the pack, not a lever for the top or bottom.

ChatGPT's citation picture is consolidating. Seer found that listicle citations dropped 30% month-over-month in early 2026. The citation set is getting narrower and more concentrated toward fewer, more authoritative sources, which means each remaining citation carries more weight and each ghost citation becomes proportionally more costly. Authoritative pages losing citations is a related pattern worth reading alongside this data.

The One Thing to Take Away

Retrieval is not citation. Citation is not visibility.

Your content clearing the retrieval threshold and your brand clearing the mention threshold are two separate systems with two separate optimization strategies. Most AI search strategies address only the first. The gap between them is where ghost citations live, and Seer's data says that gap encompasses 61.7% of all AI search appearances.

If your content is being retrieved and cited but your brand is not being named in the response, the problem is not your content. The problem is your entity. The content is doing its job. The parametric layer does not know who you are, or it knows you as something other than what you are now. That is a different problem with a different fix and a longer timeline. Entity resolution is where that work starts.