AI citation metrics

Why one number isn't enough

Most GEO measurement programs reach for a composite "AI visibility score" as the headline KPI; Profound, Peec AI, Otterly, AthenaHQ, and Ahrefs Brand Radar all surface composite-style headline scores in their public marketing. Composite scores are convenient, but they hide the structural per-engine gaps and content-property dynamics that drive citation decisions, and they collapse measurements that often disagree on the same source for the same prompt set. The cross-engine asymmetries are large enough that aggregation hides what matters: in Ahrefs' August 2025 long-tail study, the percentage of AI-cited URLs that also ranked in Google's top 10 for the same query was ~29% on Perplexity vs ~7-8% on ChatGPT / Gemini / Copilot¹. Whatever metric you choose, an aggregated cross-engine number averages signals from systems that behave structurally differently. The diagnostic information is in the per-metric, per-engine breakouts, not in the average.

GEO Glossary tracks the AI citation surface as six independent measurement dimensions, organized into three groups:

• Output ratios (three metrics): how often, against whom, and how cleanly your source is cited.
  • Attribution rate: the percentage of evaluated AI-engine responses that cite a specific source for a fixed prompt set.
  • Citation share: the relative percentage of citations a source receives versus competitors for the same prompt set.
  • Citation match rate: the percentage of AI-engine references to a source that include a clickable link back.
• Input content property (one metric): the upstream property that makes a passage extractable, quotable, and attributable.
  • Cite-ability: a practitioner-coined content property describing how suitable a passage is for AI extraction, quotation, and attribution.
• Temporal signals (two metrics): how the citation surface changes over time.
  • Citation velocity: the rate at which new AI-engine citations to a source accumulate over time; a temporal leading indicator paired with attribution rate (different units: a count per window vs a ratio), useful when attribution rate is flat but underlying citation flow is changing.
  • Citation rotation: the rate at which the sources an AI search engine cites for a given query change over time; the temporal-stability dimension.

None of the six AI-search operationalizations is defined in current vendor documentation or peer-reviewed academic literature²³. Each definition below is practitioner-coined and synthesized across GEO measurement practice. Each requires a per-engine breakout to be useful. Each has a denominator decision that needs to be locked in advance and held stable across measurement windows.

The six metrics in detail

The rest of this pillar walks each metric in turn: what it measures, the operational decisions that define it, and what it does not capture. The anchor entry for each metric is the authoritative reference for the deep definition, the verbatim formula, and the per-engine caveats; this pillar focuses on how the six fit together as a measurement framework.

Output ratios

Attribution rate: the headline KPI

Attribution rate answers the simplest measurement question: was my source cited? It is the percentage of evaluated AI-engine responses that cite a specific source or domain for a defined prompt set. The metric is conceptually related to the citation-counting and visibility measures in the 2023 Princeton GEO paper², though the per-engine and per-query-set operationalization here is the cluster's own extension, not the paper's.

The formula: (responses that cite the target source) ÷ (total evaluated responses) × 100%.

Three denominator decisions to lock in advance: per engine vs aggregated (always report per engine first; engines differ structurally and aggregation hides it), all probed responses vs only responses with at least one citation (the stricter denominator is more cross-engine comparable), and verbatim linked citation vs name mention (tighter definitions require the engine to actually link the source).

Attribution rate is the simplest entry point into the measurement framework; it returns a usable per-engine baseline within one weekly probe round of a 10-query prompt set. It is also the metric most other metrics in this cluster depend on: citation share normalizes against the same prompt set's total citations, citation match rate normalizes against the same source's own attributed references, citation velocity is its time-derivative, and citation rotation tracks the same probed source list across time.

Citation share: how big a slice of the pie

Citation share is the relative version: where attribution rate measures one source against all queries, citation share measures one source against other sources for the same queries. The formula: (target source citation instances) ÷ (total citation instances across all sources for those queries) × 100%.

Three additional axes to lock before reporting: URL vs domain vs brand level (the same entity yields three different shares depending on aggregation), deduplication rule (one count per response vs raw counts), and per-engine vs cross-engine vs traffic-weighted aggregation.

Citation share is the metric that supports competitive analysis. It is what tells you whether a topic is fragmented (no single source holds more than ~15% share, an enterable cluster) or concentrated (one domain holds 60%+ share, an entrenched canonical you should route around). It is the AI-search analog of traditional share of voice (not the direct equivalent: SEO share of voice is built on ranking impressions, citation share is built on counted citation events).

Citation match rate: the linked subset

Citation match rate counts only the linked subset of attribution: references that include a clickable URL back to the source. The formula: (linked citations) ÷ (all attributed references) × 100%.

The metric matters because linked and unlinked references have different downstream effects. A linked reference creates a direct click path; an unlinked reference contributes to brand awareness and may drive delayed effects (branded search, dark traffic, eventual direct visits) but does not create a click path. Reporting "mentioned vs cited" as a single binary loses the linked-prose-mention case where an engine names a brand inline without an attached URL.

Citation match rate effectively equals attribution rate on Google AI Overview (every source-panel entry is linked). On ChatGPT / Claude / Gemini the match rate depends primarily on whether web search or grounding is enabled per session, not on model version. Per-engine match-rate goals are the right framing; chasing one universal target collapses real engine-by-engine variation.

Input content property

Cite-ability: the upstream property the output ratios measure

Cite-ability is the only metric in the cluster that lives in the content, not in the measurement system. It is a practitioner-coined content property describing how suitable a passage is for AI extraction, quotation, and attribution. Where attribution rate / citation share / citation match rate measure the outcome of how engines treat the content, cite-ability measures the input property of the content itself.

There is no standardized cite-ability metric or scoring tool. Practitioners use a four-trait heuristic: a self-contained claim, an unambiguous subject, a sourceable assertion, and ideally memorable phrasing. The "context-free quote" test (copy a paragraph into a fresh AI chat with no surrounding context, ask the model to restate it) is the most common informal scoring method.

Cite-ability is the lever the other five output metrics react to. A program that improves attribution rate without changing cite-ability is leaning on distribution and authority, not on content. A program that improves cite-ability without seeing the output ratios move is either too early in the engine indexing cycle or working in a topic cluster already saturated by stronger sources. The two-sided signal (input property changes vs output ratio responses) is more diagnostic than either side alone.

Temporal signals

Citation velocity: the temporal leading indicator

Citation velocity is the rate at which new AI-engine citations to a source accumulate over time. The basic formula: citation velocity (period N) = new citation instances observed in period N (per fixed window length). The metric is often described as "the time-derivative of attribution rate", but the two have different units (a ratio vs a count per time) and the relationship is colloquial rather than a strict mathematical derivative. The practical claim is that velocity often moves before attribution rate does, not that one is the literal derivative of the other.

The metric inherits its shape from academic bibliometrics (Garfield 1955 onward³) but the AI-engine application is a recent extension. Several vendor blogs (UltraScout, Rankeo, Steakhouse) have published "citation velocity" definitions for AI search with looser, often single-window or competitor-relative framings; no major GEO tracking tool currently exposes velocity as a named field in a public-facing dashboard.

Velocity is the leading indicator to attribution rate's lagging indicator. A flat attribution rate over multiple weeks can mask both high-churn (new sources displacing old) and low-churn (stable repeats) realities. Velocity, computed as new-citation counts per probe round (distinct from acceleration, which is the period-to-period change in velocity), separates the two. Practitioner experience suggests attribution-rate shifts may be preceded by 2 to 6 weeks of velocity acceleration; this is a monitoring hypothesis, not a benchmark, and has not been empirically established in published research.

Three operational decisions: window length (weekly minimum because daily windows show too much noise from probe timing), novelty definition (a five-fold typology of first-seen / new-vs-previous-window / recovered / persistent / lost, not just a binary new vs persistent split), and per-engine breakout (engines have different ingestion lags; aggregate velocity hides single-engine spikes).

Citation rotation: the temporal-stability dimension

Citation rotation is the rate at which the sources an AI search engine cites for a given query change over time. In high-rotation windows, the cited-source set may change weekly or faster; in low-rotation windows, the same top sources persist for months. Rotation is the temporal-stability dimension that pairs with velocity's temporal-accumulation dimension.

The concept appears in the literature under several names. Citation volatility is the most common industry term (used for events like the widely-reported late-2025 episode in which ChatGPT's Reddit citation share fell from roughly 60% to roughly 10% over approximately six weeks, attributed by industry coverage to Google's removal of the num=100 search parameter). Source pool cycling and source rotation are alternative names; citation persistence is the inverse property. The citation-rotation entry uses the synthesis name for parallel framing with the rest of the cluster.

Three measurement disciplines: lock prompt set, cadence, and engine list before the first measurement window (rotation only shows up as signal if probes are consistent); categorize sources by persistence tier before computing rotation (top-tier authority sources often appear more persistent and dominate the share that does not rotate; mid-authority sources cycle on a weekly-to-monthly cadence in practitioner tracking); and track per-engine rotation curves over months, not weeks (single weekly snapshots show too much noise to separate rotation from random sampling variance).

The underlying driver of an observed rotation event is not always identifiable from publisher logs alone (it can involve retrieval, ranking, grounding-time selection, or UI sampling), so citation rotation is best understood as a measurement category, not a confirmed internal mechanism. No vendor has published a per-source rotation policy.

Which metric for which question

The six metrics answer different questions. A practitioner playbook for matching question to metric:

In this glossary's framing, two metrics typically anchor a GEO measurement program: attribution rate for the simple "am I cited" question, citation share for the competitive "where do I stand" question. The other four are diagnostic layers added once the basic baseline is stable. The decision is not "which one to pick" but "which one to add next" as the program matures.

What no single metric captures

Each of the six captures something the others do not, but none captures everything. Gaps that matter for GEO program decisions:

• Engine-traffic-weighted value. Citation share weighted by each engine's traffic share is the most strategically accurate variant, but reliable per-engine traffic data is often unavailable. A 22% citation share on AI Overview (with high downstream traffic) is more valuable than a 22% citation share on Copilot (with lower downstream traffic), but no metric in the cluster captures this weighting natively. Practitioners proxy this with click-through tracking when available.
• Click-through rate after citation. None of the six metrics measures whether a citation drives a click. AI citations often satisfy the user inside the answer without a referral visit. Track attribution rate (cited or not) and referral click-through rate (clicks per citation) as separate metrics; conflating them hides whether a low click-through reflects "not cited" or "cited but no click". Click-through measurement also requires solving external traffic disambiguation, which adds its own per-source measurement discipline.
• Sustained vs spike quality. Citation velocity reports a fresh count per period, but a velocity spike that decays within two weeks is not the same as a sustained level shift. The cluster does not include a "sustained quality" metric; practitioners read this from the combination of velocity (rate of new), citation rotation (persistence), and attribution rate (current state) across multiple windows.
• Cross-engine prompt-set alignment. Comparing citation share across engines assumes the same prompt set is meaningful on each surface. In practice, query reformulations differ across engines (Perplexity fan-out, AI Overview query suggestions, ChatGPT search query rewriting), and per-source citation differences can reflect engine-specific query handling rather than per-source performance. The cluster does not include a prompt-set normalization metric; this remains a known measurement caveat.
• Reactive vs organic rotation drivers. The citation-rotation entry hedges that rotation drivers are partially explainable (vendor parameter changes, partnership announcements) and partially residual noise (natural index refresh churn). No vendor has published a per-source rotation policy, and the cluster has no metric that decomposes rotation into reactive vs organic components.

Adoption sequencing playbook

For a GEO program building the measurement framework from zero, a sequenced adoption that matches the diagnostic value of each metric:

• Week 1: attribution rate per engine. Lock a 10-query prompt set, probe weekly across ChatGPT / Perplexity / Claude / Copilot / Google (with Gemini chat, AI Mode, and AI Overview each probed separately and rolled up under Google for the rate calculation). Record cited URL, domain, position in the source list, the answer's verbatim wording, and whether the citation is linked. This single decision returns a per-engine baseline by the end of the first weekly probe round.
• Continuously: citation match rate, as a free add-on. Citation match rate uses the same probe data as attribution rate; you are already recording link state per reference. Surface the match-rate breakout in the same weekly report.
• Promote to content process: cite-ability. Cite-ability is not a measurement output; it is the content-side discipline that drives the others. Add the "context-free quote" test and the four-trait checklist to the content review process. Cite-ability moves slowly; the lead indicator that the discipline is working is improvement in citation velocity (next).
• Week 9: citation share + citation velocity. By Week 9 you have eight weeks of attribution-rate data, which is enough to compute citation velocity (probe-to-probe deltas, store the per-probe citation lists not just aggregates) and to draw a competitive frame for citation share (define the topic cluster, tabulate competitor URL / domain / brand presence across the same prompt set).
• Quarterly: citation rotation. Citation rotation needs months, not weeks, of data to separate signal from noise. At the end of the first quarter, compute per-engine rotation curves over the full 12 to 13 weekly probe rounds. Use the rotation report as the quarterly diagnostic for which surfaces require more frequent monitoring and refresh versus which appear more stable once authority and source fit are established. The trade-off is observational rather than vendor-prescribed.

What to skip at every stage: composite "AI visibility scores" that bundle the six metrics into one number. They average away the per-engine, per-metric gaps that the framework exists to surface.

How the cluster fits the broader GEO surface

The six metrics in this pillar form the citation-metrics cluster within the broader GEO measurement surface. Three adjacent clusters interact with this one:

• Surface-family entries (AI Overview citation, AI Mode, Microsoft Copilot citations, Perplexity citation, Claude citation, Gemini citation, ChatGPT search citation, AI dev tool citations): each citation metric in this pillar must be computed per surface, not just per engine. The surface-family cluster describes the engine-specific measurement caveats that the per-engine breakouts in this pillar inherit.
• Brand mentions in AI answers (brand mentions in AI answers): the broader presence dimension that includes unlinked name references. Brand mention rate often exceeds attribution rate because AI engines mention well-known brands even when they do not cite them. The brand-mentions entry sits adjacent to the citation-metrics cluster but is not one of the six core anchors.
• Traffic-side measurement (external traffic disambiguation): the methodology for separating AI-driven referral visits from other traffic sources. Citation metrics measure the upstream event (was the source cited); traffic disambiguation measures the downstream event (did the citation drive a visit). Both sides of the funnel need their own measurement discipline.

The umbrella for all three clusters is generative engine optimization: the practice of writing, structuring, and signaling content for AI search engines. The citation-metrics cluster is the measurement layer of that practice; the surface-family cluster is the per-engine implementation layer; brand mentions and traffic disambiguation are the adjacent presence and outcome layers.