Calculating AI Share of Voice Across ChatGPT, Claude, and Gemini

- The AI chatbot landscape has redistributed dramatically: ChatGPT's share of combined web visits fell from 76.5% (February 2025) to 53.9% (May 2026) even as its absolute visit count held roughly flat, while Gemini rose from 5.6% to 27.9% and Claude from 1.4% to 9.2% — per Similarweb data reported by Momentic, August 2026
- Growth rates tell the competitive story more sharply than share: Gemini up 450% year over year, Claude up 855% — the fastest of the set — meaning the floor of "which platforms matter" is rising fast
- AI search visits grew an estimated 42.8% year over year (Q1 2025 to Q1 2026), climbing from 15.6 billion to 27.4 billion, per Digital Applied; roughly a third of US consumers now use an AI tool at the product-discovery stage
- The measurement gap: only 14% of marketers track AI citations, even as 43% name AI search optimization a core 2026 strategy — per Digital Applied
- Framing point: the platforms are not interchangeable surfaces; each has a distinct user base and growth trajectory, which is exactly why a single-platform reading produces a false sense of completeness
- No optimization or tactical content recommendations here — this section's job is to show that discovery is moving, and teams need a measurement system to keep up
What AI share of voice actually measures and what it does not
- Core definition: AI share of voice is the percentage of AI-generated answers that mention, cite, or recommend a brand across a defined set of category prompts, measured relative to all brand mentions in those same answers — per Digital Applied and Trakkr
- The unit of analysis is the answer, not the URL or ranking position: what matters is whether the brand surfaces inside the text a buyer reads
- Trakkr's canonical definition of the observation unit: one completed answer for one prompt × engine × market/language × run; a brand is counted at most once per answer even if its name appears multiple times
- Five distinct measures that must not be blended by default (per Trakkr):
- Mention share — answers in which your brand is present, counted once per answer, divided by all brand-answer mention units in the declared comparison set; the default competitive presence measure
- Recommendation share — answers that present your brand as a suitable option or shortlist choice, divided by recommendation units for every tracked brand; use for buyer-intent prompts only, a passing mention must not count
- Citation share — citation appearances attributed to your brand under a declared domain rule, divided by attributed citations for every tracked brand; use to compare evidence ownership
- Visibility rate — eligible answer observations in which your brand appears divided by all eligible completed answer observations including answers with no tracked brand; unlike share metrics, the denominator is answers not brand events
- Position or prominence — observed order, first-place count, or an explicitly weighted position score; keep raw positions and weights beside any weighted score
- Why blending hides the diagnosis: a brand might win a large share of citation share because its documentation is frequently sourced, yet earn a much smaller share of recommendations — averaging those erases the actionable signal (Trakkr worked example)
- What this article will calculate: mention share as the default, with the other four reported separately — this choice should be disclosed every time a number is published
- Benchmarks to anchor against once numbers are in hand: above 30% is considered strong in a competitive market; below 10% indicates significant gaps — per Rankio and Digital Applied
Why each platform is a separate visibility game, not three windows onto the same answer
- The divergence is measurable, not theoretical: ChatGPT mentions brands in 24.2% of queries at an average rank of #3.50; Gemini mentions brands in 14.1% of queries at an average rank of #1.97; only 11% of brands overlap between the two engines, with divergence highest in fragmented categories like finance and insurance — per Amicited, 2026
- Claude mentions brands in 97.3% of responses — the highest rate of any major model — per a Spotlight February 2026 analysis of over 2.4 million AI responses with 19 million citations across eight models, reported by AuthorityTech
- Each engine's source path differs:
- ChatGPT — draws on third-party review sites, comparison articles, and linked directories; tends to paraphrase without naming sources unless SearchGPT is explicitly activated; the largest citation surface but also the least traceable
- Gemini — deeply integrated with the Google ecosystem; schema markup lifts citation probability by 2.3× in Gemini answers; one brand decision in Google Search surfaces across Gmail, Workspace, YouTube Shorts, and Maps — per House of Martech and Everything-PR
- Claude — applies more deliberate reasoning, surfacing brands when supporting evidence is structured, sourced, and contextually specific to the query; users skew toward complex decision-making contexts
- Practical consequence: a brand can be recommended in ChatGPT, absent in Gemini, and described cautiously in Claude on the same day — not because one answer is wrong, but because each platform operates on a different source path, freshness window, and answer style
- A tool or methodology covering only one platform gives a false sense of completeness — and because the engines reward different content signals entirely, the gaps between them are actionable, not noise
- This section carries no optimization recommendations — the point is to motivate per-model calculation, not to prescribe what to do about the gaps
Building the prompt set that defines your denominator
- The prompt library is the most important input: a poorly constructed set produces meaningless metrics regardless of how precisely brands are counted — per Shadow
- Recommended scope: 5 to 10 prompts to start (with a plan cap of 50), per Trakkr; keep a smaller priority set (roughly 15 to 30) for regular tracking runs, per Shadow
- Five core prompt types to cover: informational, comparative, instructional, brand-specific, and transactional — per SE Ranking
- Include both branded and unbranded variations; unbranded prompts are often more revealing because they show whether AI models spontaneously surface the brand in category-level conversations — where most discovery happens
- Testing only branded prompts is one of the most common early mistakes: it tells you what AI knows about the brand, not whether AI recommends it
- Map prompts to buyer journey stages: informational queries to check citation authority, transactional queries to check brand recommendations — per SEO Clarity; a representative sample of roughly five highly focused questions per core business topic establishes a reliable baseline
- Prompt wording matters more than it appears: a 2026 preprint (arXiv:2605.27440) ran approximately 6,000 paraphrase tests and found that small wording changes produced much lower recommendation overlap than same-prompt reruns — 0.288 Jaccard similarity for cosmetic paraphrases and 0.135 for constraint-adding paraphrases, versus a 0.50–0.61 same-prompt rerun baseline
- Practical takeaway from that finding: track intent clusters, not isolated magic prompts; for important clusters, keep two or three stable variants so you can see whether the brand appears only under one phrasing or across the broader buyer need
- Concise, keyword-style prompts (e.g., "best CRM small business 2026") tend to produce more brand mentions — up to +25% average visibility — per Search Engine Journal
- Version the prompt library when prompts change; adding or removing a competitor changes the denominator, so start a new comparison series when the declared set changes — per Trakkr
The exact formula for calculating mention share, with a worked example
- Canonical formula (Trakkr): mention share = your brand-answer mention units ÷ all tracked brand-answer mention units × 100
- Denominator rule: every brand-answer mention unit earned by your brand and all declared competitors in the exact sample; brand-free answers add no mention units to the denominator, but they remain in the visibility-rate denominator
- Count a tracked brand at most once per answer observation, even if the name appears multiple times in that answer
- Per-model calculation first, then roll up:
- Run the formula separately for ChatGPT, Claude, and Gemini
- A weighted average across platforms is a secondary metric — compute it only after per-model figures are understood and clearly labeled
- Worked example (based on Trakkr's illustrative data):
- 40 eligible answers contain 60 tracked-brand mentions total
- Brand A has 18 mentions → mention share = 18 ÷ 60 × 100 = 30.0%
- Recommendation share: 6 of 24 classified recommendations → 25.0%
- Citation share: 9 of 20 attributed citations → 45.0%
- Visibility rate: 18 present answers of 40 eligible → 45.0%
- Average position: 43 total position points ÷ 18 present answers = 2.39
- Interpretation: the brand's mention share (30%) is higher than its recommendation share (25%) — it appears in the conversation but makes the shortlist less often; citation share (45%) is strongest, meaning its owned sources are frequently retrieved as evidence
- Advanced variant — weighted mention formula: analyze total word count of an answer against brand-specific word count (e.g., 50 ÷ 200 × 100 = 25% SOV) for a more granular signal than binary mentioned/not-mentioned — per ClickRank; flag this as supplementary, not a replacement for the canonical formula
- Disclose the formula chosen every time a number is published; the same underlying data produces materially different figures depending on which formula is applied — Digital Applied cites a worked example from LLM Pulse's methodology showing the same brand scoring 20% mention-based, 16.8% position-weighted, and 31.4% citation-based on the same dataset
Running prompts across all three models consistently
- Every prompt in the declared set must be sent to all three models in the same run; mixing engine sets between runs invalidates trend comparisons
- Record the following fields for every prompt × engine × run observation: completion status, brands mentioned, recommendations, source links, mention order, and the raw answer — per Trakkr's audit workflow
- Keep missing data distinct from a confirmed absence: a failed API call is not the same as a prompt that returned an answer with no brand mentions
- Minimum frequency: daily for priority prompts; AI models do not always respond the same way to the same prompt, so single-run snapshots are unreliable — per Geometrics
- A practical operating standard: at least 95% sample completion before computing share figures; investigate a five-percentage-point move only when it persists across two consecutive weekly runs — per Trakkr
- Record country, language, and account context when it can affect the answer; do not compare a US English run with a UK English run as the same observation — per Trakkr
- AI-generated responses drift meaningfully: cited domain sets shift substantially month over month in active categories; a 100–200 prompt panel run weekly is the minimum viable system, and even then the number is a reading, not a verdict — per Digital Applied
- Why this matters for Lettertrace users: Lettertrace automates the prompt-dispatch and response-aggregation steps across ChatGPT, Claude, and Gemini, storing raw outputs in a user-owned Supabase instance — no vendor holds the underlying data, and the prompt library and run schedule are fully configurable; this is the operational infrastructure the calculation described above requires
Aggregating per-model scores into a cross-platform SOV figure teams can track over time
- Per-model SOV scores should be published first and separately; the cross-platform aggregate is a secondary summary, not the primary metric
- Weighted average across platforms: compute each model's mention share, then average or weight by relative query volume or strategic priority; document the weighting scheme so future runs use the same one
- SOV is most valuable as a trend, not a snapshot: running the same prompt set weekly or monthly reveals whether GEO efforts are paying off and whether competitors are gaining ground — per Rankio
- Practical SOV benchmarks to orient the trend reading: above 30% is considered strong in a competitive market; below 10% indicates significant room for improvement; the key signal is direction — increasing SOV over time — per Rankio and Digital Applied
- What to inspect when a material move appears: drill to the prompt level and the engine level behind every shift; act when the same loss pattern repeats across multiple prompts on the same model; record any content, PR, or product change that preceded the move so future audits can attribute cause — per Trakkr's audit workflow
- Treat AI SOV as directional, not deterministic: cited domain sets drift substantially month over month; a weekly run of a 100–200 prompt panel is the minimum viable system for detecting genuine trend movement versus noise — per Digital Applied
- Store every run's raw data in a consistent schema so the denominator stays stable across periods; adding a competitor mid-series requires restarting the comparison series or clearly labeling the break
- The trend line across all three models simultaneously is the final deliverable: it shows not just the aggregate share but which platform is gaining or losing — the actionable intelligence that per-model tracking exists to produce

