AI Search Share of Voice: The Math That Actually Holds Up

Most "AI visibility scores" on the market today share one property: nobody outside the vendor that built them can reproduce the number. That's not a coincidence — it's a design choice, and it's the wrong one if you're reporting this metric to a CMO or a board. Share of voice (SOV) is a 70-year-old advertising concept with a well-defined formula. Porting it to AI search doesn't require inventing a proprietary index. It requires being precise about what you're counting, over what denominator, and per which engine.

This piece is about the calculation itself — the math, the panel design, and why per-engine reporting beats a blended composite. If you want the broader reporting and dashboarding practice, that's covered elsewhere on this site; here we're only doing the arithmetic.

Quick Answer

• If you're a CMO or board member who needs a defensible metric → report per-engine citation rates with explicit evaluation counts, because a single blended score cannot be reproduced by anyone outside the vendor that built it.
• If you're running a brand tracking program across multiple AI engines → never trust a blended AI-visibility score, because only 2.7% of domains are cited by all five major engines, making a single composite meaningless.
• If you're comparing your brand's visibility against competitors → run the identical prompt panel for named competitors and compare per-engine rates side by side, because the classic "share of a pie" framing requires an arbitrary, vendor-curated competitor set that is rarely disclosed.
• If you're designing a measurement panel with unequal run counts per engine (due to rate limits or speed) → use pooled raw counts rather than a naive average of percentages, because the naive average overweights engines with small evaluation runs and introduces bias.

Where "share of voice" actually comes from

Direct answer: In traditional media measurement, share of voice is a ratio of a brand's presence to total category presence — originally ad spend, later impressions or mentions. Wikipedia's definition states it plainly: SOV is "the percentage of media spending by a company compared to the total media expenditure for the product, service, or category in the market." A brand that spends $5 million in a $100 million category has 5% share of voice. Brandwatch's breakdown of the metric traces the same logic into digital and social listening: your mentions divided by total category mentions, expressed as a percentage.

Two things carried over cleanly from that history: (1) SOV is always a ratio against a defined universe, never a standalone score, and (2) the universe (spend pool, mention pool, category) has to be stated explicitly or the number is meaningless. Both properties are exactly what most AI-visibility dashboards drop.

The AI-search adaptation, and where it diverges from the "pie" model

Direct answer: Most vendors porting SOV to AI search keep the classic "share of a pie" framing: your citations divided by total citations across a hand-picked competitor set. The problem is that the competitor set is arbitrary, vendor-curated, and rarely disclosed — which means two tools tracking the same brand on the same day can, and do, produce different numbers. There's a cleaner and more auditable formulation that avoids the arbitrary-competitor-set problem entirely:

Share of voice = citation events won ÷ total evaluations run

An "evaluation" is one execution of one prompt against one engine on one occasion. A "citation event" is won when your brand is named, linked, or sourced in that specific response, using a fixed rule you write down in advance (mention-only vs. domain-link vs. both are three different — and equally valid — definitions, but you have to pick one and keep it constant).

This is a rate, not a market-share pie slice. It doesn't require you to decide who your five competitors are before you can report a number, and it's trivially auditable: anyone can take your prompt panel, run it against the same engine, and check whether your brand shows up. You still get competitive context — just by running the identical panel for named competitors and comparing rates side by side, rather than forcing everyone into one blended percentage that has to sum to 100.

Why a single blended score isn't a metric — it's a marketing claim

The complaint that AI-visibility scores are unauditable isn't a fringe take. One review of GEO dashboards found that "the same brand produces different scores across different GEO platforms on the same day," and traced the cause directly to non-disclosure: none of the major platforms publish the arithmetic behind their headline number, so "third party trust in the vendor's methodology alone is not verification" (source). The same piece cites a 2026 decision-confidence survey in which only 49% of marketing and finance leaders could clearly explain their own measurement approach to their board, and 74% had scaled back initiatives specifically because of measurement confidence gaps.

The academic literature on generative-engine evaluation backs this up from a different angle. A recent position paper on GEO governance argues explicitly for "deployment-aligned metrics" and "black-box auditing of material influence" as correctives to exactly this opacity problem (arXiv:2606.12439). The original GEO research from Princeton, which introduced citation-impact scoring for generative engines, is instructive for a different reason: it computes its citation metric by sampling the model's output multiple times per query and then explicitly normalizing against a baseline, precisely because a single response is not a stable measurement (arXiv:2311.09735). If the researchers who invented citation scoring for generative engines don't trust a single sample, a vendor dashboard that reports one snapshot number shouldn't be trusted either.

Per-engine, not blended: a worked example

Direct answer: Different AI engines cite completely differently. One comparative analysis of citation behavior found ChatGPT averaging roughly 3.7 sources per answer against Perplexity's 8.6 and Gemini's 11.0, with citation overlap so low that only about 2.7% of domains got cited by all five major engines studied (source). That alone should kill the idea of a single universal AI-visibility number — an engine that cites sparsely and an engine that cites liberally are not comparable on a blended scale without badly distorting one or the other.

Here's what a defensible panel run looks like, using the citation-events-won-over-evaluations formula, with unequal run counts per engine (a common real-world constraint — some engines rate-limit or are slower to query):

Report those four rows. That's the metric. If a single blended figure is unavoidable — say, for an executive summary — there are two ways to compute it, and they are not interchangeable:

• Naive average of the four percentages: (23.3 + 50.0 + 7.5 + 24.4) ÷ 4 = 26.3%
• Pooled raw counts (correct if you must blend): (28 + 30 + 9 + 22) ÷ (120 + 60 + 120 + 90) = 89 ÷ 390 = 22.8%

The naive average overweights Perplexity's small 60-evaluation run purely because it happened to have a high hit rate, pulling the blended figure up by 3.5 points versus the pooled calculation. This is the exact failure mode that makes single composite scores untrustworthy even when the underlying data collection was honest — the blending method itself introduces bias if evaluation counts differ across engines, which they almost always do in practice.

How many evaluations you actually need

Direct answer: Large language models are non-deterministic by construction — identical prompts can produce different citations on different runs, a property documented directly in empirical studies of model output variance (arXiv:2308.02828). That means a single run per prompt per engine isn't a measurement, it's a coin flip you're reporting as fact.

The nearest disciplined analogy is opinion polling, not classic rank tracking. Political and market-research polling has a century of literature on why sample size and repeated measurement determine how much you can trust a percentage: AAPOR's guidance on margin of sampling error notes that precision gains taper sharply once you're past a few hundred observations, and that margin-of-error math only applies cleanly to random or representative samples in the first place (AAPOR; Qualtrics). Applied to AI-search SOV: running each prompt once per engine per week is a convenience sample, not a probability sample, so don't report a formal confidence interval on it — but you should still run each prompt multiple times per measurement window (industry guidance clusters around 30-50 evaluations per engine as a floor before the per-engine rate is treated as stable) and disclose that repetition count next to every number you publish.

A minimal, auditable methodology

Direct answer: If you're building or evaluating a share-of-voice number for AI search, it needs to survive these checks:

1. The prompt panel is published or at least disclosed in count and composition — buyer-intent questions a real prospect would ask, not branded queries about your own product.
2. The citation-event rule is fixed in advance — mention, domain link, or both, applied identically to every evaluation.
3. Every number is denominated against total evaluations run, and that N is stated alongside the percentage.
4. Results are reported per engine first, with any blended figure computed from pooled raw counts, never averaged percentages.
5. The underlying evaluation log is available for spot-checking — someone should be able to take five prompts from your panel, run them by hand against the same engine, and land in the same neighborhood you reported.

This is the model behind nqzai's answer-engine visibility tracking: it runs a documented prompt panel independently against each major AI engine, logs every citation event against a fixed rule set with a timestamp and the source response, computes per-engine rates before any blending, and keeps the raw evaluation log available so a reported number can be checked rather than taken on faith.

The takeaway

Direct answer: Share of voice in AI search isn't a new metric that needs a new, proprietary formula. It's the old advertising ratio applied to a new medium, with one adjustment: because the medium is non-deterministic and fragmented across engines with wildly different citation behavior, the calculation has to be a rate over a disclosed panel, reported per engine, not a single composite number nobody can check. If a vendor hands you one score and won't show you the panel, the run count, or the citation rule behind it, you're not looking at a measurement — you're looking at a claim.

Sources:

If you want to see this measurement approach applied to your own brand rather than building the tracking yourself, nqzai's AI Share of Voice tool runs the same query-sampling methodology automatically.

• Share of voice — Wikipedia
• What Is Share of Voice? Formula & Examples — Brandwatch
• Measuring AI Share of Voice Across LLM Answer Engines
• Share of Voice in AI Search: How to Calculate It
• AI Visibility Dashboards Grade Their Own Homework
• Margin of Sampling Error/Credibility Interval — AAPOR
• Margin of Error Guide — Qualtrics
• GEO: Generative Engine Optimization — arXiv:2311.09735
• An Empirical Study of the Non-determinism of ChatGPT in Code Generation — arXiv:2308.02828
• Position: Generative Engine Optimization Creates Underexamined Risks — arXiv:2606.12439