Citation share is the fraction of AI-composed answers, across a fixed set of buyer questions, that cite or link your site as a source; share of model (also called AI share of voice) is the broader sibling: the fraction of those answers that mention or recommend your brand at all, cited or otherwise. Together with per-engine variants and sentiment reads, these form the metric family that makes AI visibility a number you can baseline, trend and be held accountable for. This page is the canonical set of definitions and the measurement discipline that keeps the numbers honest.
The metric family, defined precisely
| Metric | Definition | What it tells you |
|---|---|---|
| Citation share | Of the answers to your fixed question set, the fraction citing or linking your site, per engine | Whether engines treat you as a source worth showing |
| Share of model / AI share of voice | The fraction of those answers mentioning or recommending your brand, cited or uncited | Whether engines treat you as an answer to the question |
| LLM citations (count) | The raw tally of your pages cited, and which pages, across the answer set | Which assets are earning selection, and where gaps sit |
| Competitive citation share | The same fractions computed for each named competitor on the same set | Whose answers these currently are, and what is displaceable |
| Sentiment / framing | How mentions characterize you: recommended, listed, caveated, warned against | Whether presence is helping or hurting |
Two distinctions keep the family coherent. Citation and mention are different events: an assistant can recommend you by name without linking anything (a mention that builds no referral path), or cite your page in support of a generic claim without featuring you (a citation with little brand effect). Track both, because they respond to different work - citations follow source selection, mentions follow entity strength and corroboration. And every metric here is conditional on the question set: none of them is "your visibility" in the abstract, all of them are "your visibility on these questions." That is a feature - it ties the number to questions with commercial stakes - but only if the set is treated as a fixed instrument. This family is the AI-era extension of classic share of voice measurement, with the panel of engines standing in for the media channels.
How to measure honestly
The method is simple to state and easy to corrupt. Fix the question set: the real questions your buyers ask - definitional, comparative, "best X for Y," problem-shaped - written down, versioned, and changed only deliberately, because every edit resets comparability. Ask on a schedule: weekly is the working cadence; engines refresh sources and models at their own pace, and a schedule is what turns snapshots into a trend. Ask per engine: ChatGPT, Perplexity, Gemini and Copilot select differently, and Google's AI features in Search behave differently again, so a blended number hides exactly the variation you need to see. Archive the answers: store the full text with dates, because the archive is what makes movement provable, regressions diagnosable, and framing shifts visible. Then compute the fractions per engine and in aggregate, for you and for each competitor. Expect variance: assistants are stochastic, so treat single-run flips as noise and multi-week trends as signal. This loop is the measurement spine of the AI visibility audit, run as a cadence rather than a one-off.
What moves the number
Citation share responds to the levers of answer engine optimization, and it helps to name them in the metric's terms. Retrievability gates everything: an engine cites nothing it cannot fetch and parse, so crawler access, clean structure and structured data come first. Liftable formats win the selection: plainly stated definitions, honest comparison tables and answers that begin at the beginning are what engines quote. Corroboration moves mentions most: facts about you appearing consistently across independent sources read as consensus, and consensus is what a model repeats. Topical depth compounds: clusters that cover a subject beat isolated pages, because engines favor sources with demonstrated authority on the question's topic. The research behind generative engine optimization - the original GEO paper - found measurable visibility gains from exactly these content-side interventions, which is the academic version of what practitioners see: the metric moves when the underlying material becomes easier to select.
Reading the number without fooling yourself
Three disciplines keep the metric useful. Read it against the baseline and against named competitors: absolute citation share varies with how contested the question set is, so "we moved from a fifth to a third of answers while the leader held flat" is a claim, and a bare percentage is decoration. Route findings into work: every uncited question is either a content gap, a retrievability bug or a corroboration gap, and the metric's job is to say which. And watch losses as seriously as gains - a competitor displacing you inside answers you used to hold is the earliest warning AI visibility gives, and the archive is what lets you catch it the week it happens. On AstroFabric this whole loop ships as a standing program: the AI Visibility agent runs the scheduled question set, per-engine tracking and answer archive on cadence, and the AI visibility solution maps the metrics into the broader growth, revenue and operations picture.
Frequently asked questions
What is citation share?
The fraction of AI-composed answers, across a fixed set of target questions, that cite or link your site as a source - computed per engine, on a schedule, against an archived baseline. It measures whether answer engines treat you as a source worth showing.
What is share of model?
The fraction of answers to your question set that mention or recommend your brand at all, cited or uncited - also called AI share of voice. It is the mention-level sibling of citation share and responds most to entity strength and corroboration across independent sources.
How do you track LLM citations?
Fix a versioned question set, put it to each engine on a schedule, archive the full answers with dates, and record which pages are cited and which brands are mentioned. The fractions per engine, trended against a baseline and competitors, are the tracking.
What is a good citation share?
There is no universal benchmark - the number depends on how contested your question set is and how many sources engines cite per answer. The honest standards are relative: above your baseline, trending up, and gaining against the competitors measured on the same set.
Why does citation share fluctuate week to week?
Assistants are stochastic and engines refresh models and sources continuously, so individual answers vary between runs. Scheduled measurement with archived answers separates that noise from real movement: treat single-run flips lightly and multi-week trends as the signal.
Sources
- Google Search Central - AI features and your website
- GEO: Generative Engine Optimization (the original research paper)
- llms.txt - the proposed standard for LLM crawler guidance
Every playbook on this blog ships as a runnable mission.
Open a workspace and the playbook library is waiting - describe the outcome and the agents carry it end to end, on your plan's monthly credits.