Report four metrics separately — citation rate, share of voice, stability, cross-platform concordance — each with its sample, window, engine and citation definition attached. Never collapse them into a score out of 100. Follow that and any two measurements become comparable; skip it and "visibility" means whatever the reporter wants it to mean.
- Each metric below is written to be quoted verbatim: formula, unit, population and collection method in one block, so a definition can be lifted without the surrounding argument.
- A properly reported figure names four things a score out of 100 hides — how many queries, on which engine, over what window, and what counted as a citation.
- AI answers are non-deterministic, so a single-run measurement conflates signal with noise. This standard asks for three to five runs per query and the spread reported alongside the mean.
- The standard deliberately excludes a composite visibility score. The weighting between inputs is the substantive claim, and it is the part vendors do not disclose.
- This is a proposal, not an adopted standard. No vendor is bound to it, and most current tools do not follow it.
The four metrics, defined
Each block below is self-contained. Take one, quote it, and it carries its own formula, unit, population and collection method with no reference back to this page.
Citation Rate (CR)
Definition. The share of queries in a fixed set for which a given domain is cited at least once by a named engine.
Formula. CR = (queries in which the domain was cited) ÷ (total queries in the set).
Unit. A percentage, 0–100%.
Population. One named, frozen query set, measured on one named engine, within a stated collection window. Never a blend of engines.
Collection method. Every query run at least three times within the window. State in advance whether a query counts as cited when the domain appears in any run or in a majority of runs; the two produce different numbers on the same data.
What it solves. It is the base unit, and no two tools currently define it the same way.
AI Share of Voice (SOV)
Definition. A domain's citations as a proportion of all citations awarded across a fixed query set by a named engine.
Formula. SOV = (citations of the domain) ÷ (all citations of all domains in the same set and window).
Unit. A percentage, 0–100%.
Population. One named query set on one named engine. The denominator is engine-specific, so two SOV figures from two engines are not on the same scale.
Collection method. Count every cited domain in every run, not only your own. Report the composition of the denominator alongside the share, because a rise can mean a rival fell.
What it solves. It makes competitive comparison possible on a fixed denominator, which citation rate alone cannot.
Citation Stability
Definition. How consistently the same query returns the same cited sources across repeated identical runs.
Formula. For each query, the share of runs in which the domain appeared; stability for the set is the mean of those per-query shares across queries where the domain appeared at least once.
Unit. A proportion between 0 and 1.
Population. The same frozen query set, same engine, same window as the citation rate it accompanies.
Collection method. Three to five identical runs per query, issued from a controlled session — logged out, location fixed, no query history — because a measurement taken from a working browser is a measurement of that browser.
What it solves. It quantifies how much of a reported number is noise. A period-over-period movement smaller than your own run-to-run spread is not a result.
Cross-Platform Concordance
Definition. How much two engines agree on which domains to cite for the same query set.
Formula. A Jaccard index — the intersection of the two engines' cited-domain sets divided by their union.
Unit. A coefficient between 0 and 1; 1.0 is perfect agreement, 0 is no overlap at all.
Population. One query set, run on exactly two named engines, in the same window.
Collection method. Both engines sampled over the same dates with the same citation definition. A definition that differs between the two engines makes the coefficient meaningless.
What it solves. It prevents a result on one engine being reported as a result about AI search.
Concordance is in the set because published overlap figures are low and inconsistent.
Ahrefs reports limited overlap between engines; Seer found 87% of SearchGPT citations matching Bing's top results; and Search Engine Journal reports AI Overview citations drifting away from top-ranking pages. Three studies, three surfaces, three definitions. The gap between Google's own two surfaces is measured in the source delta study.
How to report a figure
Under this standard a citation-rate claim reads: "Domain X had a 34% citation rate across 200 queries on ChatGPT Search, collected across 5 runs each between [dates], with a stability score of 0.71." Not: "Domain X scored 82/100 on AI visibility." The first can be checked, disputed or replicated. The second cannot.
| Before (typical vendor phrasing) | After (standard-compliant) |
|---|---|
| "Your AI visibility score improved from 61 to 74 this quarter." | "Citation rate across our 150-query brand set rose from 22% to 31% on ChatGPT Search (5 runs each, stability 0.68 → 0.74); Perplexity citation rate was flat at 18%. AI Overviews were not included in this quarter's measurement." |
The compliant version is longer and less quotable as a headline. That trade-off is deliberate: it tells the reader exactly what improved, on which engine, with what confidence, and what was not measured at all. Applied to this site's own work, it is what the AI Citation Index reports and what the dataset strategy stores.
The five principles
- 1 Define citation Link, named mention, or paraphrase. Written down before any counting starts.
- 2 Run repeatedly Three to five runs per query. Report the spread, not only the mean.
- 3 State sample and population How many queries, which engines, what window. Without this, nothing is falsifiable.
- 4 Label the claim Correlational or causal. Almost everything in this field is the former.
- 5 Publish components Citation rate, share of voice and stability separately. Never a single blended score.
Why the term is broken today
AI answers are non-deterministic. The same query can return a different set of cited sources on a second attempt, which is why concordance has to be measured across repeated runs rather than inferred from one. Almost every published AI-visibility figure comes from a single run.
"Citation" is not consistently defined. Whether an unlinked brand mention counts, or a paraphrase without attribution, is answered differently by different tools and usually not disclosed. The unlinked-mention question is not academic: including or excluding those mentions changes a citation rate by itself.
Sample and population are usually undisclosed. A visibility score rarely states how many queries it rests on, across which engines, over what window — making it unfalsifiable by construction. The same weakness runs through the figure set traced in the provenance audit.
The underlying reason is structural: every major player measuring this has a commercial reason to define it favourably. Even the careful vendor studies — Semrush on AI Overviews and Ahrefs on citations and the top 10 — sit on corpora nobody outside the company can inspect. Terms used here without definition are in the glossary.
Building a compliant query set
Every principle depends on the query set, and the query set is where most measurement quietly fails. A biased set produces disclosed, checkable, useless numbers.
Start from demand, not from your content. Queries chosen because you already rank for them guarantee a flattering result, and a set built from pages you already have will keep rediscovering the domains that already dominate.
Include queries you expect to lose. A set with no failures is not measuring anything. Losses are where change becomes visible first.
Cover the intent range deliberately. Definitional, comparison, how-to, purchase. Intent matters because engines fan a single question out into several sub-searches, as the granted patent describes and the fan-out corpus study measures.
Fix the exact wording and freeze it. Paraphrases are different queries. A set that drifts cannot support a trend claim.
Hold back a control group. Queries relevant to your category that you are not working on. When everything moves together, they tell you the platform moved, not you.
Publish the set if you can. A published query set is the single strongest disclosure available. It lets someone else run your measurement and disagree with your conclusion using your own instrument.
Computing stability, with arithmetic
The numbers here are invented for illustration. They are not measurements.
Run one query five times and your domain is cited in three of them: for that query you were cited 60% of the time. Run a second query five times and it is cited in all five. A single-run tool records both as "cited". They are not the same result, and the difference is exactly what stability captures.
That is the practical payoff of principle two. It converts "did it change" from a comparison of two numbers into a comparison against noise. Without it, every measurement programme reports movement every period, because something always moves.
Seven pitfalls that break comparability
One: changing the query set between periods. The most common and most damaging. A comparison across two different sets is not a trend.
Two: silently changing the citation definition. Counting unlinked brand mentions this quarter and not last quarter produces a rise on its own, and correlational ranking-factor work suggests that is a large swing rather than a rounding difference.
Three: mixing engines into one figure. A blended rate hides which engine moved, and engines behave very differently — see this side-by-side of how each platform cites.
Four: counting domain instead of URL, or the reverse. Both are defensible. Switching between them mid-programme is not.
Five: uncontrolled personalisation. Signed-in sessions, saved location, prior queries. Google's documentation on AI features is the only first-party account of what varies, and it does not describe personalisation in enough detail to control for it.
Six: sampling at a fixed time of day only. If system behaviour varies with load, a fixed sampling time bakes that variation into your trend.
Seven: reporting the mean without the spread. A mean alone cannot be interpreted. This is principle two restated, because it is the one most often skipped under deadline.
What does not transfer between engines
| Does not transfer | Why not |
|---|---|
| Citation rate | Engines differ in how many sources they cite at all. A higher rate can be an interface property, not a visibility gain. |
| Stability | Some engines are far more repeatable than others. Comparing raw stability across engines compares their architectures. |
| Share of voice | The denominator is engine-specific. Two share figures from two engines are not on the same scale. |
| The effect of a change | A page edit visible in one engine within days may take much longer to register in another, or never — the question the crawl-to-citation latency study exists to measure. |
| What counts as a citation | Inline links, source lists and unlinked mentions appear in different proportions per engine. |
The standard's answer is not to normalise these away. It is to report per engine, always, and to treat any cross-engine aggregate as a summary rather than a measurement.
Why no composite score
The exclusion of a composite score is the most contested part of this proposal. It rests on four points.
The weights are the claim, and they are hidden. A score combining citation rate, share of voice and sentiment embeds a theory about how much each matters, and that theory is exactly the part not disclosed. Compensation hides direction: a score can hold steady while citation rate falls and something else rises.
It cannot be replicated. Every component in this standard can be recomputed by an outsider from disclosed inputs; a proprietary blend cannot. And it becomes a target — a score clients are managed against will be optimised, including by the vendor who defines it. This is the same objection the tactic evidence scoreboard makes to blended tactic rankings.
We expect composite scores to persist regardless of this argument, because they are commercially useful. The realistic goal is disclosure of components alongside the score, not the score's disappearance.
Confounds in period-over-period reporting
Even a fully compliant measurement can support a wrong conclusion.
The platform changed, not you. Engines ship changes without announcement. Control queries are the only cheap defence, and almost nobody keeps them.
Competitors changed. Share of voice is relative; a rise can mean a rival got worse rather than you got better.
Seasonality in the query set. Some questions are asked differently at different times of year, so a quarter-over-quarter comparison can be a comparison of seasons.
Your own unrelated changes. A migration, a redesign or a content pruning will move these numbers. Keep a change log alongside the measurement log or attribution becomes guesswork.
Regression to the mean. A period measured after an unusually bad one tends to look better regardless of action taken. This is the confound most likely to make an intervention look effective when it was not.
The clearest published example of what a compliant negative result looks like is the llms.txt evidence: an adoption study, a request-tracking study, and a 300,000-domain analysis finding no clear effect. This site's own version is a first-party log study.
A minimum viable version
Full compliance costs real money. A stripped version is far better than nothing, and it is honest about being stripped.
Thirty queries, frozen, five of them controls. Small, but enough to see large movements and to catch platform-wide shifts.
Three runs per query, not five. Three is enough to distinguish "always" from "sometimes", which is most of the value.
One engine, named. Measuring one engine properly beats measuring four badly. Check first that the engine can reach you at all — access gates everything, which is what the bot registry and the technical audit cover.
One definition of citation, written down. Whatever you choose is fine as long as it is fixed and stated.
One sentence of disclosure attached to every figure. Sample, window, engine, definition. That single sentence is the standard's core.
This version is cheap enough for a small team to run monthly. It is not as good as the full specification, and saying so is part of complying with it.
What would move this to v2
- A shared public query setHighest impact
Cross-vendor comparability
The largest remaining barrier is that everyone measures different questions. A common open set would do more than any of the five principles.
- Evidence on stability thresholdsAwaiting data
Principle 2
We decline to publish a threshold because we have no basis for one. Data on typical run-to-run spread would let a real threshold replace judgement.
- Operator sampling interfacesNot offered
All engines
If engines exposed a controlled sampling interface, most of the sampling difficulty here would disappear.
- Evidence a component is redundantOpen
Metric set
If concordance carried no information beyond the per-engine rates, we would drop it. A standard that only ever grows is one nobody is testing.
Limitations
- This is a proposal, not an adopted standard. No vendor is bound to it, and most current tools do not follow it.
- Perfect compliance is expensive. Multi-run sampling across several engines costs more than a single-run check; the standard states the ideal, not a requirement for every casual use case.
- It does not resolve the definition of "citation" itself — it requires you to state your definition, not adopt a specific one, because reasonable definitions vary by use case.
- Disclosure alone does not prevent a poorly designed sample from producing a misleading, if technically checkable, result.
- Citation rate is an intermediate metric. Whether it produces revenue is a separate question this standard does not address — see conversion benchmarks.
- Not every surface can be sampled repeatably. Where principle two is unachievable, say so rather than reporting a single-run number as compliant.
Last verified: September 2026
- Rewrote the four metrics as self-contained definition blocks carrying formula, unit, population and collection method, so each can be quoted without this page.
- Cut the media-measurement analogy, the objections section and the stakeholder-reporting section; none carried evidence.
- Moved the metric definitions above the diagnosis of why the term is broken.
- v1 proposal published: five principles, four metrics, no composite score.
Next: run the minimum viable version on your own site. Start by checking which questions you currently appear on with the AI Overview checker, then freeze that list as your query set. To see the full specification applied end to end, read the AI Citation Index.
Namdev, R. (2026). The AI Visibility Measurement Standard (v3). Retrieved from https://ritiknamdev.com/blog/ai-visibility-measurement-standard Published under CC BY 4.0 — reuse freely with attribution.
This is the measurement pillar under AI SEO. It is applied directly in the AI Citation Index's methodology, referenced by the GEO tactic evidence scoreboard's grading system, and used to report the entries in the null-results registry.