The five AI visibility metrics that matter: (1) citation rate — how often your brand is cited in answers to your money questions; (2) share of voice — your citations versus a named competitor set in the same answers; (3) the retrieval-visibility-to-citation-visibility gap — being fetched versus being cited; (4) AI referral volume — sessions that arrive from AI surfaces, in GA4; and (5) generative-AI impressions — appearances in Google Search’s AI features, in Search Console. Each covers a different stage of the answer pipeline, and each has a hard limit on what it can tell you; the limit matters as much as the definition.
This is the measurement post of the content & corroboration layer in our AEO/GEO basics series — the one-page overview has the reading order. The setup for the official tools in detail is in how to measure AI search traffic; the report shape these numbers live in is the one-page monthly template; and the levers that move the citation-side numbers — extractability and entity clarity — are in what makes content AI-citable and entity clarity for AI answers.
Why five, not one
An answer to a question goes: retrieve → select → generate with citations → (sometimes) click. No single number covers more than one stage, and the official dashboards each cover at most one stage of one surface:
- Citation rate and share of voice cover the citation stage, per question and per surface — visible only by running the questions yourself, because no official tool scores chat-assistant answers.
- The retrieval/citation gap covers the retrieve and select stages.
- AI referrals cover the click stage, and only the tail of it that survives to your site.
- Generative-AI impressions cover the generate stage — on Google Search’s own features only.
That is why a program reports all five: a brand can be retrieved but uncited, cited but not mentioned favorably, mentioned but never clicked, and all of that while its impression count climbs. One number will always say “fine” while another is failing.
The five, defined
1. Citation rate
Definition: across a mapped set of money questions, how often your brand is cited (with a link), mentioned (without one), or absent — per surface, per month. How measured: by running the question set across the surfaces — ChatGPT, Perplexity, Gemini, Claude — on a schedule, capturing answers verbatim, and scoring each. The standing version is a harness; the one-off version is an hour of manual checking (the per-surface procedure is in the measurement guide). What it cannot tell you: why an answer is shaped that way; whether the mention is favorable (that is scored from the captured answer text, not from the rate itself); or what one surface does — surfaces retrieve and cite differently, so a single-surface rate is a partial number by construction (the surface comparison).
2. Share of voice, versus a named competitor set
Definition: in the same answers, how your citation count compares to the specific competitors the answers name — not a market estimate, but a named set, per question set. How measured: the same probes, scored against the named set; the answers themselves usually name who the alternatives are, so the set stays honest. What it cannot tell you: absolute market size (you are measuring share of cited presence in your question set, not share of the market), and the denominator problem: most citations in your category’s answers go to review platforms, community and comparison sources, not to vendors at all (why answers triangulate). A 100% share among named vendors can coexist with a small presence in the answer as a whole.
3. Retrieval visibility vs citation visibility
Definition: the gap between being retrieved — fetched into the engine’s candidate pool — and being cited in the final answer. The 2026 observational study of production ChatGPT web search is the evidence that these are separate quantities: “candidates retrieved more often are also more likely to be cited,” and the study’s stated implication is that GEO should “distinguish retrieval visibility from citation visibility” (arXiv:2610.15382, https://arxiv.org/abs/2610.15382/, observational, not causal). How measured: citation side by the probes above; retrieval side by proxies — what Google Search and the API-level search candidates return for your questions. One hard caveat from the same study: those proxies combined “recover only about one quarter of cited URLs,” so the proxy view is a rough map of the pool, not the pool. What it cannot tell you: the selection logic between retrieval and citation — unpublished in every engine — or the full candidate set, which is largely private.
4. AI referral volume (GA4)
Definition: sessions that arrive at your site with an AI-surface referrer — chatgpt.com, perplexity.ai, claude.ai, gemini.google.com and the relevant Copilot domains. How measured: GA4 referral rows, segmented by referrer domain — deliberately not by UTM, because surfaces frequently strip or drop your campaign parameters (the full setup, including the drop caveats, is in the measurement guide). What it cannot tell you: citations — many citations produce no click at all, so high citations with low referrals is a normal, healthy pattern — and it is a floor, not a census: some referrals arrive with the referrer stripped and land in (none)/(direct). Read it as direction month over month, not level.
5. Generative-AI impressions (GSC)
Definition: per Google’s own definition, “how many times links to your site were shown to a user in a generative AI feature on Google Search” — the report, launched June 3, 2026 and rolled out worldwide August 31, 2026, covers AI Overviews and AI Mode (Google Search Central: https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports; https://support.google.com/webmasters/answer/16984139). How measured: it is a report you open, not a probe you run — monthly totals and the per-page split, which is your AI-feature exposure map. The scoping note that bites most: Bing’s equivalent AI Performance report (public preview since February 10, 2026, https://www.bing.com/webmasters/help/ai-performance-9f8e7d6c) counts Bing’s own AI features — it is not a ChatGPT report, and neither report says anything about chat assistants. What it cannot tell you: clicks, positions, or citations; anything outside Google Search’s features; and single-month moves deserve care, because Google changes which queries get AI features independently of your content. The full interpretation, including why three dashboards will disagree with each other, is in Google’s generative AI report, decoded.
What the set looks like in a month (SAMPLE)
The table below is the shape of the citation-side metrics (1 and 2) as they appear in the monthly report. This is SAMPLE data — illustrative rows, not live results for any real brand — the same shape as the sample scorecard in that post:
| Money query | ChatGPT | Perplexity | Gemini | Claude | Δ vs last month |
|---|---|---|---|---|---|
| “best [category] for [use case]” | Cited #2 | Absent | Cited #4 | Mentioned, no link | ▲ improved |
| “[your product] vs [competitor]” | Absent | Cited #1 | Cited #2 | Cited #3 | — flat |
| “[competitor] alternatives” | Cited #3 | Cited #2 | Absent | Absent | ▼ declined — fix queued |
| “[product] pricing” | Cited #1 | Absent | Cited #2 | Cited #1 | ▲ improved |
The other three metrics sit alongside it as one row each — AI referral sessions per surface, and GSC generative-AI impressions — with their month-over-month deltas. Five numbers, five stages, one page. The rules for what never goes in that page (no traffic promises, no unattributed numbers, sample labeled sample) are in the report template post.
The rules that keep the set honest
- Impressions are not citations, citations are not clicks, clicks are not a census. Each metric stops where its source stops. Cross-comparing them produces the false conclusions the decoded GSC report post walks through.
- Compare each metric against itself, month over month. Windows, feature lists and model versions all drift; a number compared across months on its own definition is the only comparison that means something.
- Label the sample. Anything illustrative is labeled SAMPLE; anything measured names its source, date and query set. That is the honest-data policy the whole site runs on, and it is documented in the methodology.
Getting the first real set
The five definitions above are the measurement; the first instance of it is a snapshot — one run of your money questions across ChatGPT, Perplexity, Gemini and Claude, with the citation-side numbers: cited, mentioned, or absent, per surface, with what each answer actually said. GA4 and GSC you read yourself in an hour each, per the measurement guide. After that, the number that matters is the delta — and the standing version of the whole run is in the methodology, with the series overview in AEO and GEO in one page.