Share of AI Voice

A Formal Framework for Quantifying Brand Visibility in AI Assistant Answers

Version 2.0•October 2026

Abstract

We define Share of AI Voice (SAV), the metric Xaudient uses to measure how often AI assistants name a brand when buyers ask them questions, and its two companions, WSAV and Organic SAV. As AI assistants increasingly mediate how buyers discover products, media and search metrics no longer see the answers buyers read. SAV is measured on the answers of four AI assistants (ChatGPT, Claude, Gemini and Perplexity) to a confirmed set of buyer questions, against confirmed competitors.

1. Introduction

Unlike search engines that return ranked lists of URLs, AI assistants synthesize information into conversational answers that name a few companies. Brand visibility in this context depends on whether a brand is named, where in the answer it appears, and in what Tone.

Definition 1.1 (SAV). SAV is your brand's mentions as a share of all mentions across your confirmed competitive set.

Definition 1.2 (WSAV). WSAV (weighted share of AI voice) is the same share weighted by position in the answer and by tone.

Definition 1.3 (Organic SAV). Organic SAV covers queries that name nobody: it is a true share (every company's share of all mentions in that bucket; brand plus competitors sum to ~100%).

2. Mathematical Framework

2.1 Fundamental Definitions

Let us establish the formal notation:

Q = {q₁, q₂, ..., qₙ} — Set of confirmed questions
P = {p₁, p₂, ..., pₘ} — Set of AI assistants (ChatGPT, Claude, Gemini, Perplexity)
B = {b₁, b₂, ..., bₖ} — Your brand and its confirmed competitors
R(q,p) — Answer from assistant p to question q

2.2 Mentions

m: R × B → ℕ

where m(r, b) is the number of mentions of brand b extracted from answer r.

2.3 Position

The position of a mention is its order among the brands named in the answer: the first brand named has position 1. WSAV weights each mention by position:

Position 1 → 1.0
Position 2 → 0.8
Position 3 → 0.6
Position 4 → 0.4
Position 5 or later → 0.3
Mentioned, no position → 0.2

2.4 Tone

Each mention carries a Tone, and WSAV weights it:

Positive → 1.2
Neutral → 1.0
Mixed → 0.8
Negative → 0.5

English Monitoring Cycles use a keyword-based Tone indicator. Italian Monitoring Cycles carry a Provisional Tone: AI-assessed, not yet human-validated. Other languages report Tone as not measured, so WSAV is not measured for them.

3. SAV, WSAV and Organic SAV

3.1 SAV

SAV is your brand's mentions as a share of all mentions across your confirmed competitive set.

SAV(b) = Σ_(q∈Q) Σ_(p∈P) m(R(q,p), b) / Σ_(c∈B) Σ_(q∈Q) Σ_(p∈P) m(R(q,p), c) × 100%

SAV = brand mentions ÷ (brand mentions + confirmed competitors' mentions) × 100%

3.2 WSAV

WSAV (weighted share of AI voice) is the same share weighted by position in the answer and by tone.

WSAV = Σ brand (position weight × Tone weight) ÷ Σ all (position weight × Tone weight) × 100%

Where Tone is Provisional, WSAV is Provisional too.

3.3 Organic SAV

Organic SAV covers queries that name nobody: it is a true share (every company's share of all mentions in that bucket; brand plus competitors sum to ~100%).

Organic SAV(b) = SAV computed over Q_organic ⊆ Q

4. Reading SAV Against Competitors

Every company in B is counted the same way: your brand gets its SAV, and each confirmed competitor is shown by its number of mentions. Only Organic SAV shows every company's share. With n companies, an even share is 100% ÷ n: in a market of four, 25% is parity, and above it a brand is named more often than its competitors on average.

5. Query Stratification

SAV varies significantly across question types. We partition the question space:

Question category: Market, Competitor comparisons, Pain-driven, Buyer journey, Geography
Question bucket: Open-market, Head-to-head, Brand-context

Stratified analysis reveals where brands excel or underperform:

SAV(b | category) = SAV computed over Q_category ⊆ Q

6. Temporal Dynamics and Trend Analysis

SAV is not static. AI assistants are retrained and updated, and brand perceptions evolve:

SAV(b, t) = SAV computed in Monitoring Cycle t

The rate of change provides insight into momentum:

ΔSAV(b, t) = SAV(b, t) - SAV(b, t-1)

Deltas compare consecutive Monitoring Cycles of the same series, and never cross a change of language.

7. Applications and Use Cases

7.1 Brand Positioning Strategy

Identify query categories where SAV lags competitors, informing content strategy and partnership decisions.

7.2 Product Marketing Optimization

Analyze which product features earn first-position mentions versus late or unranked ones.

7.3 Competitive Intelligence

Track competitor SAV trends to detect emerging threats or market share shifts before they manifest in traditional metrics.

7.4 Campaign Attribution

Measure SAV changes following major campaigns, product launches, or PR events to quantify impact on AI-mediated brand perception.

8. Limitations and Future Work

While SAV provides valuable insights, several limitations warrant acknowledgment:

Query Coverage: Comprehensive query generation remains challenging; observed SAV depends on query sampling strategy.
Provider Variability: LLM responses exhibit non-determinism; multiple samples per query-provider pair may be necessary.
Tone: Automated Tone is imperfect; Italian Tone is Provisional until human-validated, and WSAV carries the same label.
Causality: SAV measures correlation, not causation; attribution to specific marketing actions requires controlled experiments.

9. Conclusion

Share of AI Voice represents a paradigm shift in brand visibility measurement, addressing the unique characteristics of LLM-mediated information discovery. As conversational AI becomes increasingly central to user experiences, SAV provides the quantitative foundation for strategic decision-making in this new landscape.

Future research directions include multi-modal SAV (incorporating image and video responses) and causal inference frameworks linking marketing actions to SAV changes.

References

Brown et al. (2020). "Language Models are Few-Shot Learners." NeurIPS.
Anthropic (2024). "Claude 3 Model Card." Technical Report.
OpenAI (2023). "GPT-4 Technical Report." arXiv:2303.08774.
Google (2024). "Gemini: A Family of Highly Capable Multimodal Models." Technical Report.
Aaker, D. (1996). "Measuring Brand Equity Across Products and Markets." California Management Review.