AI Visibility Metrics Dictionary: 18 Terms With Formulas
AI visibility measurement has a vocabulary problem: different tools use the same words for different calculations. The answer is not to pretend a universal standard exists. It does not. This is an operational dictionary—a compact, reproducible set of definitions for teams measuring whether a brand appears, is recommended, and is supported by sources in AI answers.
Table of contents
- How to read the formulas
- The 18-metric dictionary
- Worked example
- Minimum reporting standard
- Limitations
- Frequently asked questions
How to read the formulas
Let N be the number of valid prompt-engine observations in a reporting window. An observation is one prompt sent to one engine at a recorded time. Define a brand mention, rank, citation, and conversion before collecting results; otherwise, the denominator moves while the report is being written. AEOeye's published methodology is one concrete example of making these choices explicit.AEOeye methodology
The 18-metric dictionary
The table is designed to be copied into a measurement plan. “Count” means count observations, not unique prompts, unless the formula says otherwise.
| Metric | Exact operational definition | Formula |
|---|---|---|
| Mention rate | Share of observations in which the brand is named or clearly identified in the answer. | mentions / N × 100 |
| Citation rate | Share of observations containing at least one attributable citation or source link that supports the answer. | cited observations / N × 100 |
| Share of model | The brand's share of all observed brand mentions in a defined prompt set and engine. | brand mentions / all tracked-brand mentions × 100 |
| Share of voice | The brand's share of recommendation appearances, counting each brand once per observation to avoid long lists dominating. | observations recommending brand / observations recommending any tracked brand × 100 |
| Average rank | Mean ordinal position when the brand appears in a ranked list; unmentioned observations are excluded. | Σ observed ranks / mentioned observations |
| Weighted visibility | A position-adjusted presence score where earlier recommendations receive more weight. Publish the weight function. | Σ position weight(r) × sentiment weight / N × 100 |
| Sentiment distribution | The percentage of mentioned observations classified as positive, neutral, or negative under a fixed rubric. | sentiment class count / mentioned observations × 100 |
| Prompt coverage | The share of distinct tracked prompts where the brand appears at least once across selected engines. | prompts with ≥1 mention / distinct prompts × 100 |
| Engine coverage | The share of tested engines that mention the brand at least once in the window. | engines with ≥1 mention / engines tested × 100 |
| Competitor win rate | Among head-to-head observations, the share where the brand is recommended and the named competitor is not, or ranks lower under the pre-set rule. | brand wins / head-to-head observations × 100 |
| Source diversity | The number of unique source domains (or a normalized diversity index) cited across observations. State which version you use. | unique cited domains (or 1 − Σ share(domain)²) |
| Citation correctness | The share of audited citations that actually support the claim they are attached to. | supporting citations / citations audited × 100 |
| Citation completeness | The share of material, externally checkable claims that have an adequate citation. | material claims with adequate citation / material claims audited × 100 |
| Answer volatility | How often the measured result changes across repeated runs of the same prompt and engine. | changed runs / repeat runs × 100 |
| Confidence interval | A range expressing sampling uncertainty around a rate; report the method and confidence level. | For a proportion, p ± z × √(p(1−p)/N) (approximate 95% when z=1.96) |
| Sample size | The number of valid observations contributing to a metric after exclusions and retries. | N = valid prompt × engine observations |
| Observation window | The fixed time span during which observations are collected, including timezone and run schedule. | start timestamp → end timestamp |
| Conversion rate | The share of defined AI-attributed sessions or users completing a chosen conversion event. | conversions / AI-attributed sessions (or users) × 100 |
“Citation” should mean an inspectable source, not merely a confident sentence. FActScore evaluates factuality at the atomic-claim level, while ALCE focuses on generating answers with citations; those are useful anchors for correctness and completeness, but neither paper creates an industry-wide business KPI.FActScore ALCE
Worked example
The following values are explicitly hypothetical. Suppose a team tests 40 prompts across 4 engines, producing N = 160 observations during one week. The brand appears in 52 answers, receives a supporting citation in 31, and appears in 35 distinct prompts. It is recommended in 70 observations where any tracked brand is recommended; tracked brands receive 140 total recommendation appearances under the one-per-observation rule.
- Mention rate:
52 / 160 × 100 = 32.5%. - Citation rate:
31 / 160 × 100 = 19.4%. - Prompt coverage:
35 / 40 × 100 = 87.5%. - Share of model: if all tracked-brand mentions total 208, then
52 / 208 × 100 = 25%. - Share of voice:
70 / 140 × 100 = 50%.
For a rough 95% interval around the mention rate, p=.325 and N=160 gives approximately 32.5% ± 7.2 percentage points, or about 25.3%–39.7%. This approximation assumes independent observations; repeated prompts, correlated engines, and changing model behavior can make the effective sample smaller. Report the calculation as a guide, not a guarantee.
Minimum reporting standard
A credible dashboard should show the metric definition beside the number. At minimum, publish:
- The exact prompt list or a versioned sampling description, including buyer intent and locale.
- Engines, model/version when available, run timestamps, timezone, and observation window.
- Sample size, exclusions, retries, and whether each denominator is prompt-, engine-, answer-, session-, or user-based.
- The rank, sentiment, citation-support rubric, and competitor tie-breaking rules.
- Point estimates plus uncertainty or a plainly stated reason uncertainty is unavailable.
- Raw answer captures or stable hashes so another reviewer can audit classifications.
For conversion rate, keep AI visibility data separate from analytics attribution. Google Analytics documents traffic-source dimensions such as source, medium, and campaign; map those fields to a clearly named AI source rather than claiming every direct visit was caused by an AI answer. Google Analytics dimensions
Limitations
AI answers are stochastic and may vary by account, location, personalization, tool settings, retrieval state, and time. A single run is a snapshot. Volatility is itself a result worth reporting, not noise to silently discard.
Metrics also encode judgment. “Mentioned,” “supports,” “positive,” and “wins” need annotation rules and periodic quality checks. A high share of voice can coexist with poor factual accuracy; a high citation rate can coexist with weak or outdated sources. NIST's AI Risk Management Framework emphasizes measuring, documenting, and managing context-specific risks—use that mindset when deciding what a metric may and may not support.NIST AI RMF
Finally, do not compare scores from different tools unless their prompts, engines, definitions, weights, and windows are aligned. Trend your own series first. If you want a transparent starting dataset, run an AEOeye audit and inspect the underlying prompt-level evidence.
Frequently asked questions
Is share of model the same as share of voice?
No. Share of model counts brand mentions; share of voice counts recommendation appearances under a stated rule. Keep both when answers contain discussion that is not an endorsement.
Should unmentioned answers count in average rank?
Usually no: average rank describes placement conditional on appearing. Report mention rate beside it so a brand with one top placement cannot look equivalent to a consistently visible brand.
What makes a citation “correct”?
The cited page must support the specific claim, in the relevant context, at the time audited. A related domain or a page that merely repeats the keyword is not enough.
How often should metrics be refreshed?
Use a fixed cadence appropriate to decision speed—weekly for active experiments is a reasonable starting point—and preserve each window so changes can be compared. Re-run after major model, site, or campaign changes.
Measurement becomes useful when its definitions are boringly clear. Save the prompt set, publish the denominator, show the uncertainty, and let the trend—not a mysterious composite score—guide the next AEO decision. Run a free AEOeye audit to turn this dictionary into a prompt-level baseline.
FAQ
Is there a universal standard for AI visibility metrics?+
No. There is no universal industry standard for the terms or weights in AI visibility reporting. This page is an operational dictionary: a transparent set of definitions AEOeye uses so teams can reproduce and compare their own measurements over time.
How many prompts do I need to measure AI visibility?+
There is no magic number. Use a deliberately sampled set of buyer-intent prompts, report the sample size and observation window, and attach uncertainty to rates. More prompts help only when they represent real intents rather than repeated paraphrases.
What is the difference between mention rate and citation rate?+
Mention rate asks whether the brand appears in an answer. Citation rate asks whether a source is linked or otherwise cited for the answer, usually with the brand or its supporting page as the source. A brand can be mentioned without receiving a citation.
Can AI visibility metrics predict sales?+
They are leading indicators, not sales forecasts. Conversion rate connects measured AI referrals to outcomes, but attribution is incomplete when users switch devices, search directly, or never click a cited source. Use visibility metrics alongside analytics and qualified pipeline data.
Sources
Is AI recommending you?
Run a free AI visibility audit and find out in under a minute.