Five Citation Patterns: Stable, Volatile, Carousel, Decay, Ghost

If you run the same query through an AI engine ten times, you do not get the same answer ten times, and you do not get the same citations. Because answers are generated rather than retrieved from a fixed index, citation behavior has texture: a source can be cited every single time, or half the time, or rotated with rivals, or cited today and gone next month. These are not noise. They are patterns with causes, and the pattern you are in determines what intervention will actually move you.

The mistake is treating citation as a single state, cited or not. That hides the most actionable information you have. A source cited 50% of the time and a source cited 50% of the time can be in completely different situations, one volatile and fixable through stability work, one in active decay and needing fresh signal. The five patterns below are how you tell them apart.

Stable: cited every run

Stable is the goal. The source appears in the answer consistently across repeated runs of the same query, which means the engine treats it as a dependable, high-confidence source for that query. Stable citations are bankable: you can forecast traffic and revenue from them because they reliably reach users. There is no fix needed here, only protection. The risk to a stable citation is complacency, because the other patterns are where a stable position erodes if you stop maintaining the content. Identify your stable citations and treat them as assets to defend, not problems to solve.

Volatile: cited unpredictably

Volatile means the source appears some runs and not others, with no clear pattern. This is the engine being genuinely uncertain about whether you belong in the answer. You are close to the threshold: good enough to be cited sometimes, not strong enough to be cited reliably. Volatility usually traces to extractability or authority consistency, the content is borderline quotable, or your authority signal is inconsistent enough that the engine includes you only when its retrieval happens to surface your stronger signals.

The fix for volatile is to push past the threshold. Strengthen the extractability of the specific passages that should be getting cited, make them cleaner and more self-contained, and tighten your authority consistency so the engine encounters the same strong signal every time rather than a mix. Volatile is the most encouraging pattern to be in, because it means you are already winning sometimes; the work is converting "sometimes" into "always," which is usually a smaller lift than getting cited in the first place.

Carousel: rotated with a fixed set of competitors

Carousel is a specific kind of volatility where you and a fixed set of competitors are rotated through the same citation slot, each appearing some fraction of the time. The engine sees you and your rivals as roughly interchangeable for this query, so it cycles among you. This is different from plain volatility because the cause is not your weakness but your sameness: you are not differentiated enough from the rotation set for the engine to consistently prefer you.

The fix for carousel is differentiation, not strength. Adding more of the same signals your competitors also have keeps you in the rotation; what breaks you out is covering an angle or a facet they do not, so the engine has a reason to prefer you specifically for some slice of the query. This connects to information gain: the engine rotates interchangeable sources, so the way out is to be non-interchangeable, to say something the others do not. Carousel is a positioning problem wearing a visibility costume.

Decay: cited at first, then fading

Decay is a temporal pattern: the source is cited strongly at first, then progressively less over weeks or months. This is the natural aging of content in a system that weights recency. Your content was fresh and relevant, it earned citations, and as it ages and competitors publish newer material, the engine gradually prefers the fresher sources. Decay is not a failure of the original content; it is the half-life of an unrefreshed asset in a recency-aware system.

The fix for decay is refresh and re-establish recency. Update the content with current information, update the dateModified, add new material that re-signals freshness, and the engine re-evaluates you as current rather than aging. The key is catching decay early, because a citation in decay is still reaching some users, and refreshing before it fully fades is far easier than recovering a citation you have already lost entirely. This is why citation stability has to be monitored over time, not measured once.

Ghost: informs the answer but is never shown

Ghost is the cruelest pattern. Your content shapes the answer, the engine clearly used your information, but you are never shown as a visible citation. The user benefits from your work and never learns your name. This happens at the surfacing layer: your content was good enough to retrieve and use, but your source authority was not high enough for the engine to display you as a citation it stands behind, so it used your information while citing a more authoritative source for the same claim, or no source at all.

The fix for ghost is authority, specifically the kind that earns visible attribution. The engine is willing to read you but not to put its name next to yours publicly, which is a trust judgment about your domain. Raising your source authority, through the external corroboration and entity strength that signal trustworthiness, is what converts a ghost into a visible citation. Ghost is the hardest pattern to even detect, because it does not show up as a citation at all; you find it by noticing the answer contains your specific framing or facts while crediting someone else.

Why the pattern determines the fix

The throughline is that "increase my citations" is not an actionable instruction until you know your pattern, because each pattern responds to a different lever. Volatile needs extractability and consistency. Carousel needs differentiation. Decay needs refresh. Ghost needs authority. Apply the carousel fix to a decay problem and nothing moves, because you are differentiating content the engine is ignoring for being stale. The diagnostic, running the same query repeatedly and watching the pattern, is what tells you which lever to pull.

The contrarian point is that citation volatility is information, not noise to be averaged away. Most measurement collapses repeated runs into a single percentage and throws away the shape, which is exactly the part that tells you what to do. Run your key queries multiple times, classify the pattern for each source, and you will know precisely which intervention each citation needs. This pattern analysis sits on top of the four-layer way an LLM picks sources, since ghost is a surfacing-layer problem and volatile is usually an attribution-layer one, and it feeds directly into how you track share of model over time rather than as a single snapshot.

Sources

• Princeton, GEO: Generative Engine Optimization: research on citation behavior in generative engines. arxiv.org/abs/2311.09735
• OpenAI, ChatGPT search: how non-deterministic generation produces variable citations. help.openai.com
• Perplexity, on source ranking: how the answer engine selects and rotates cited sources. perplexity.ai/hub/blog
• Website AI Score, the Citation Stack: the layered model that explains where each pattern originates. View article
• Website AI Score, share of model vs rank tracking: tracking citation patterns over time. View article