AI Share of Voice — measure who AI cites for your category.

Classic share of voice measures your slice of a finite, observable space: ad impressions, media mentions, search rankings. AI Share of Voice measures something stranger — your slice of the answers a generative engine composes on the fly. There is no fixed inventory to divide; each answer is synthesized anew, so SOV is the probability, sampled over many answers, that the engine chooses to mention and cite you.

The first distinction from classic market share is the nature of the medium. Market share counts transactions that already happened; AI SOV counts the upstream moment when an assistant shapes what a buyer even considers. Because a growing share of buyers now ask an engine before they decide, presence in those answers is a leading indicator of the market share that follows, not a lagging report of it.

The second distinction is presence versus citation. An engine can mention your brand without crediting you as a source, or cite you as the authority behind a claim. These are different signals: a mention puts you in the consideration set; a citation marks you as the trusted origin. Rigorous SOV measures both, because a brand that is mentioned but never cited is visible yet not authoritative.

The third distinction is the probabilistic nature of generative answers. Ask the same question twice and an engine may answer differently, citing different sources each time. That variance is not noise to be ignored; it is the phenomenon itself. A single answer is an anecdote. SOV only becomes a metric when you sample the same question repeatedly and aggregate, turning a stochastic process into a stable, trendable number.

This is why AI SOV cannot be read off a single prompt. One answer on one engine on one day tells you almost nothing, because the next answer may differ. Treating a lucky or unlucky single response as a measurement is the most common error in early generative-visibility work, and it produces conclusions that reverse the moment you sample again.

It also differs from a search ranking. A ranking is an ordered list the same algorithm produces consistently; an AI answer is a fresh synthesis that may include or omit you for reasons that shift with phrasing, context and the engine's current grounding. SOV measures inclusion in that synthesis, which is both more valuable — you are in the answer, not tenth on a page — and harder to measure, because it must be sampled rather than simply read.

The unit of analysis is the query set, not the keyword. Because engines answer questions, SOV is defined over a deliberately chosen set of questions a buyer in your category would actually ask. The composition of that set is a methodological choice with real consequences: too narrow and you measure a corner of the market; too broad and you dilute the signal. A defensible SOV starts with a defensible query set.

Engine and time window are part of the definition, not afterthoughts. The same brand can have very different SOV on ChatGPT versus Perplexity, and SOV drifts as engines update and the web changes. A SOV figure is therefore always qualified by which engines, which queries and which window it covers; an unqualified percentage is not interpretable, and comparing two such numbers measured differently is meaningless.

SOV is best understood as a distribution, not a point. Across your query set and engines, you are cited often on some questions, never on others, and intermittently on the rest. The headline percentage summarizes that distribution, but the shape matters: broad, shallow presence and narrow, deep authority are different strategic positions that the same average can hide, which is why SOV is read alongside the pyramid and the per-engine breakdown.

Crucially, SOV is causal-friendly when measured well. Because you can attribute a movement to a specific change — a published asset, an earned reference, a corrected entity — SOV closes the loop between AEO and GEO work and its effect. A measurement system that cannot attribute movement reduces optimization to guesswork; one that can turns SOV into the steering signal for the entire generative-visibility program.

It is also inherently comparative. Your SOV only means something against the category — the competitors the engine could have cited instead. A SOV of ten percent is strong in a crowded category and weak in a sparse one. Rigorous measurement therefore tracks not just your share but the shape of the competitive field around it, so a leadership team can see whether a gain came from growing the category presence or from taking share from a named rival.

Finally, SOV is the metric that makes the whole method accountable. AEO and GEO are inputs; SOV is the outcome they are optimizing. By defining it precisely — presence and citation, over a fixed query set, on named engines, in a stated window, sampled enough times to be stable — you turn generative visibility from a topic of opinion into a number a board can track quarter over quarter and a team can be held to.

There is a subtlety worth naming: SOV measures answers, but answers are produced for people, so the metric is ultimately a proxy for influence over buyer consideration. A brand cited in the answer a buyer reads has shaped that buyer's shortlist before any classic funnel metric registers anything. SOV is therefore not a vanity number adjacent to the funnel; it is a measurement of the funnel's new, AI-mediated front door.

It follows that SOV deserves the same governance as a financial metric. A number a board steers by must be defined once, computed identically every period, and qualified by its conditions, exactly as revenue is recognized under a fixed policy. Treating SOV casually — different queries each quarter, single prompts, no engine breakdown — produces a figure that looks like a metric but cannot bear the weight of a real decision.

Seen this way, AI Share of Voice is the natural successor to the share-of-voice metric marketers have tracked for decades — relocated from media and search to the answers generative engines compose. The questions it answers are the same a CMO has always asked: are we present where buyers decide, are we credible there, and are we gaining or losing ground against named rivals? Only the surface being measured has changed.

A single headline percentage hides too much, so SOV is decomposed into four KPIs that build on one another. Together they distinguish a brand that merely appears from one that is cited, cited prominently, and cited consistently everywhere — four very different strategic positions a single average would blur into one.

The first KPI is SOV presence: across your query set, in what share of answers does your brand appear at all? Presence is the entry condition — if you are absent from the answer, nothing downstream matters. It answers the most basic question, do we show up, and it is the floor on which the other three KPIs are built.

The second KPI is the citation tracker: of the answers where you appear, how often are you cited as a source rather than merely mentioned in passing? Volume of citation separates being named from being credited. A high presence with low citation means the engine talks about you but trusts others as the authority — a signal to invest in the GEO work that earns source status.

The third KPI is the weighted citation score: not all citations are equal, so this KPI weights each by its prominence in the answer. A citation in the opening sentence carries more influence than one buried at the end, just as the first result historically outweighed the tenth. Weighting by position turns a raw count into a measure of how much your citations actually shape the answer the buyer reads.

The fourth KPI is the cross-platform consistency score: does your message hold across ChatGPT, Claude, Perplexity and Gemini, or only on one? A strong signal on a single engine is fragile; a coherent presence across all four is robust authority. This KPI captures the weak-but-consistent signals across engines that, taken together, indicate durable rather than incidental visibility.

Read as a set, the four KPIs form a diagnostic. Low presence points to an extractability or coverage gap — an AEO problem. High presence but low citation points to a trust gap — a GEO problem. Good citation but poor weighting points to prominence you can improve with answer-first structure. Inconsistency across engines points to a coherence gap in your entity. Each KPI routes you to a specific remedy.

The KpiQuad below shows the four at a glance, but their power is in their relationships. Presence and citation together tell you whether visibility is converting into authority; weighted score and consistency together tell you whether that authority is influential and durable. Watching the four move relative to one another is far more informative than tracking any one in isolation.

They also map cleanly onto the optimization levers. Presence and citation respond to AEO and GEO work on the questions where you are weak; weighted score responds to answer-first prominence; consistency responds to entity consolidation across surfaces. Because each KPI has a corresponding lever, the four-KPI view is not just a scoreboard but a work-allocation tool that tells you where the next effort will pay off most.

Importantly, the four KPIs are measured the same way every time. Each is computed over the same defined query set, the same engines and the same sampling protocol, so movements reflect real change rather than methodological drift. A KPI that is computed differently from one period to the next cannot be trended, which is why the rigour of the underlying protocol matters as much as the choice of KPIs.

Rolled together, the four KPIs produce the single SOV figure a leadership team tracks, but the decomposition is what makes that figure actionable. The headline tells you whether you are winning; the four KPIs tell you why, and where to act. That combination — one number to steer by, four to diagnose with — is what turns SOV from a vanity metric into an operated control system.

A final caution: the four KPIs must not be collapsed prematurely into one. The temptation to report a single, tidy SOV percentage is strong, but the decomposition is precisely what makes the metric diagnostic. A leadership team can absolutely steer by the headline, provided the four underlying KPIs remain available beneath it, because the day the headline moves the first question is always which of the four moved, and why.

In practice, the four KPIs also give the program its language. Conversations stop being about whether AI likes the brand and become precise: presence is up but citation is flat, weighting improved on grounded answers, consistency dipped on one engine after a profile change. That precision is what lets a team plan the next sprint against evidence rather than impression, and it is the quiet payoff of decomposing SOV rather than reporting a lone percentage.

The pyramid is a way to read the four KPIs as a hierarchy of difficulty and value. Each level is harder to reach and rarer than the one below it, so the shape narrows as you climb: many answers mention you, fewer cite you, fewer still cite you prominently, and fewest of all do so consistently across every engine. The geometry encodes the strategy.

The base is presence: the broad foundation of answers in which your brand appears at all. It is the widest level because appearing is the easiest of the four to achieve and the precondition for everything above. A narrow base caps the whole pyramid — you cannot be cited prominently and consistently in answers you never appear in.

The second level is citation volume: of the answers where you are present, those that credit you as a source. This level is narrower because being cited is harder than being mentioned; it requires the trust that GEO builds. Moving mass from the presence level up into the citation level is one of the clearest signs that authority work is paying off.

The third level is citation quality: the citations that land prominently enough to shape the answer. Narrower again, because prominence is scarcer than mere inclusion. This level reflects answer-first structure and strong entity signals working together, and it is where citations stop being a footnote and start steering what the buyer concludes.

The apex is cross-platform weak signals: the consistent, corroborating presence across ChatGPT, Claude, Perplexity and Gemini that, taken together, indicates durable authority. It is the rarest level because coherence across four independently-grounded engines is the hardest thing to achieve, and the most valuable because it is the least dependent on any single engine's quirks.

Reading the pyramid top to bottom diagnoses your position. A pyramid that is wide at the base but pinched immediately above signals presence without authority. One that is healthy through quality but collapses at the apex signals single-engine strength that is fragile. The shape of your pyramid, more than any single number, tells you which level is the binding constraint on your visibility.

The pyramid also disciplines goal-setting. Rather than chasing a single SOV percentage, you target movement of mass up the levels: more presence converted to citation, more citation made prominent, more prominence made consistent. Each upward shift is a concrete, measurable objective tied to a specific lever, which is far more actionable than a vague ambition to raise the headline figure.

Finally, the pyramid is the bridge between the four KPIs and the rigorous protocol that measures them. The levels only mean something if each is measured the same way every time, over the same query set and engines, with enough sampling to be stable. The next section describes that protocol — the position-weighted formula, the N=5 sampling, the A/B/C classification and the consistency score that make the pyramid trustworthy.

The pyramid metaphor also guards against a common misreading of averages. Two brands can share an identical headline SOV while having opposite pyramids — one broad and shallow, present everywhere but cited nowhere prominent, the other narrow and tall, cited deeply on a few questions. They face entirely different next moves, and only the level-by-level view, not the single percentage, makes that difference visible and therefore actionable.

The first pillar is the position-weighted citation, or PWC, adapted from the academic GEO literature. Instead of counting citations equally, PWC weights each by its rank in the answer, because a citation in the opening line influences the reader far more than one at the end. The formula sums each citation multiplied by a rank weight that decays with position, so prominence, not just presence, drives the score.

Position weighting matters because generative answers are read top-down and often truncated. A source named first frames the entire answer; a source named last may never be read at all. Treating those as equal would overstate the influence of buried citations and understate the brands that own the opening — exactly the distinction PWC exists to capture, and the reason a raw citation count is a misleading SOV proxy.

The second pillar is N=5 multi-sampling: each query is asked at least five times, and the results are aggregated. Because generative answers vary from one sampling to the next, a single response cannot estimate the true citation probability; repeated sampling does. Five is the practical floor at which the average stabilizes enough to trend reliably without an impractical measurement cost.

Multi-sampling is what converts variance from a problem into information. The spread across the five samples is itself a signal: a brand cited in all five is robustly present, while one cited in two of five is marginal and volatile. Averaging gives the central estimate; the dispersion around it tells you how reliable that estimate is, which a single sample can never reveal.

The third pillar is A/B/C answer classification, which sorts each answer by how the query was grounded. Class A answers draw on the engine's parametric memory, class B on web-grounded retrieval, and class C on product-adjacent or branded framing. The same brand can have very different SOV across these classes, so classifying answers prevents averaging together fundamentally different measurement conditions.

The classification matters because parametric and grounded answers are governed by different levers. Parametric SOV reflects what the model absorbed in training — slow to move, GEO-driven. Grounded SOV reflects what the model retrieves now — faster to move, AEO-driven. Reporting them separately tells you which lever is working; collapsing them into one figure hides the mechanism and makes optimization a guessing game.

The fourth pillar is the consistency-and-authority score, or CAS, which captures how coherent and how well-sourced your presence is across engines and samples. It rewards a brand that is cited consistently and from credible sources, and discounts presence that is erratic or thinly grounded. CAS is what distinguishes durable authority from a lucky spike, and it is the score that best predicts whether a SOV gain will persist.

Together these four pillars form a measurement discipline rather than a single trick. PWC handles prominence, N=5 handles variance, A/B/C handles grounding mode, and CAS handles coherence. Remove any one and the metric degrades: drop position weighting and you overcount footnotes; drop sampling and you measure noise; drop classification and you blend incompatible conditions; drop CAS and you mistake spikes for authority.

The point of the rigour is not academic. It is what lets a leadership team trust the number enough to act on it and a practitioner attribute a movement to a specific change. A SOV figure produced without position weighting, sampling, classification and a consistency score is not wrong by a little — it can be wrong by enough to send a program in the opposite direction from where it should go.

None of this requires gratuitous jargon to use. The team consuming SOV needs to know that citations are weighted by prominence, that every query is sampled multiple times, that parametric and grounded answers are reported apart, and that consistency is scored — not the underlying mathematics. The rigour lives in the protocol so the readout can stay plain, which is exactly how a trustworthy metric should be built.

It is worth stressing that these four pillars are interdependent, not a menu. You cannot apply position weighting meaningfully without first sampling enough to know which citations are stable; you cannot classify A/B/C usefully without enough samples per class; and the consistency score is only interpretable once the other three are in place. The protocol is a system in which each pillar makes the others trustworthy.

Above all, the rigour is what earns SOV its place on a dashboard next to revenue and pipeline. A metric leadership acts on must be defensible under scrutiny, and a SOV built on position weighting, multi-sampling, grounding classification and a consistency score can be explained, audited and trusted. That defensibility is not overhead; it is the difference between a number a board steers by and a chart no one quite believes.

The four major engines ground their answers on overlapping but distinct sources and update on different schedules, so being cited by one is no guarantee of being cited by the others. Cross-platform measurement asks a sharper question than per-engine SOV: not just whether you are present somewhere, but whether your presence is coherent everywhere it matters.

A spike on one engine is fragile for a simple reason — it depends on that engine's particular grounding, which can change with the next model update or web shift. A brand whose visibility rests on a single engine is one update away from a cliff. Consistency across four engines is far harder to dislodge, because it would take simultaneous shifts across independently-grounded systems to erase it.

The consistency matrix below maps your key claims against the four engines, marking where each claim holds and where it diverges. A claim that is consistent across all four is robust authority you can rely on; a claim that holds on two and fails on two is a coherence gap — usually a sign that your entity or your evidence reads differently to different engines and needs reconciling.

Divergence is diagnostic, not just a defect. When an engine tells a different story about your brand, it is reporting what its particular sources say, so a divergence points you at the specific surface — a stale profile, a missing reference, an inconsistent description — that is feeding that engine a different picture. Fixing the divergence fixes the underlying inconsistency, not just the symptom on one engine.

Consistency also compounds trust within each engine. Engines increasingly cross-check claims, and a brand whose message is the same everywhere reads as more credible than one whose story shifts by platform. Coherence is therefore not only a robustness property across engines but an authority signal inside each one, which is why cross-platform consistency sits at the apex of the SOV pyramid.

Measuring four engines is also what keeps SOV honest. Optimizing for a single engine invites overfitting — tuning to one system's quirks at the expense of the others. Tracking all four forces you to build authority that generalizes, the kind that rests on genuine source quality and entity coherence rather than on a trick that happens to please one engine this quarter.

Coverage extends beyond the four as new prescriptive engines gain traction, but the principle is constant: the more independent surfaces on which your message holds, the more durable your authority. A consistency score is, in effect, a measure of how diversified and therefore how resilient your generative visibility is against the inevitable churn of any single engine.

Read together with the other KPIs, consistency is the durability check. Presence, citation and weighting tell you how strong your visibility is right now; consistency tells you how likely it is to last. A leadership team watching SOV should read consistency as the risk indicator — high consistency means the gains are banked, low consistency means they could evaporate with the next model update.

One practical consequence is that consistency reframes prioritization. When a claim holds on three engines and fails on one, the highest-leverage move is usually to fix the failing engine's underlying source rather than to push harder on the three where you already win. Consistency measurement thus continually redirects effort toward the specific gap that is dragging your durable authority down, which a single blended SOV number would never surface.

Run them in order and keep them fixed: the value of SOV comes from measuring the same way every time, so the discipline is as much about not changing the protocol between periods as about following it once. A measurement that drifts methodologically cannot be trended, and an untrendable SOV is back to being an anecdote.

Define the perimeter before you measure anything. The query set, the engines, the time window and the sampling count are all part of the measurement, and changing any of them silently invalidates comparison with prior periods. Pin them down first, document them, and treat the perimeter itself as version-controlled so a number always carries the conditions under which it was produced.

The cadence is part of the rigour. Engines re-synthesize continuously and the web changes, so a one-time measurement is a snapshot that ages immediately. Running the protocol on a fixed rhythm — the same query set, engines and sampling each period — is what turns SOV from a snapshot into a trend, and only a trend can show whether your authority is rising, flat or quietly eroding.

Attribution is what makes the trend useful. Because the perimeter is held constant, a movement in SOV can be traced to the specific change that caused it — a published asset, an earned reference, a corrected entity. That attribution is the feedback that lets a program double down on what works and stop what does not, which is the entire point of measuring SOV in the first place.

Done consistently, this protocol turns SOV into the steering signal for the whole method. AEO widens the questions you can answer, GEO deepens the trust behind those answers, and the five-step measurement loop tells you, period over period, whether the combined effort is moving the number that matters. Without the protocol, the method has no feedback; with it, every AEO and GEO investment is judged against an honest, attributable measure of generative authority.