AI Share of Voice: The Formula, a Worked Example, and Real Benchmarks

Share of voice, as a concept, has existed in marketing for decades: your brand's mentions divided by total category mentions, across ads, social, or search. What's new is applying that same division to a source that didn't exist five years ago, AI-generated answers, and the formula changes just enough that treating "AI Share of Voice" as a drop-in replacement for classic SOV gets the number wrong.

We built our AI Share of Voice tracking around a real, published dataset: 240 AI-engine responses (ChatGPT, Claude, Gemini, and Perplexity, in English, Portuguese, and Spanish, 20 buying-intent prompts per engine per language) in our own AI-visibility-tools category. This piece walks through the exact formula, shows the full worked calculation using that dataset, and gives you the benchmarks and the step-by-step process to run the same measurement for your own brand.

What AI Share of Voice Actually Measures

AI Share of Voice (AI SOV) is the percentage of AI-engine answers, across a defined prompt set and a defined set of engines, that cite a given brand. It answers a narrower, more specific question than "is my brand visible in AI search": it tells you, out of every time someone in your category asks an AI engine a buying-intent question, how often you're the one who gets named.

Two things distinguish it from classic multi-channel share of voice. First, the "impressions" side of the calculation isn't ad spend or search volume, it's a fixed set of real, repeatable prompts run through real engines, because AI answers aren't queryable as a stable inventory the way ad auctions or keyword volumes are. Second, a "mention" in an AI answer can be detected two different ways (a cited source domain, or the brand name appearing in the generated text), and the two signals have different reliability, a distinction that matters enough to get its own section below.

The Formula

At its simplest:

AI Share of Voice = (Responses citing your brand ÷ Total responses in the measurement set) × 100

That's the whole formula. The complexity isn't in the math, it's in defining the measurement set correctly: which prompts, which engines, which languages, and how long a citation window counts as a "response." Get the denominator wrong (too few prompts, too narrow an engine set, prompts that don't reflect real buyer language) and the percentage that comes out the other end is precise-looking but meaningless.

A second, closely related metric worth calculating alongside AI SOV is relative share, your citations divided by the category leader's citations, rather than by the total. If you're at 8% and the leader is at 33%, your relative share is roughly 24% of the leader's presence, a number that often communicates the size of the gap to a stakeholder more directly than the raw percentage does on its own.

A Full Worked Example, Using Real Data

Here's the calculation applied to our own category, using the actual numbers from our July 2026 measurement.

The measurement set: 20 buying-intent prompts (questions like "what are the best AI visibility tools" and "what tools track my brand's visibility in ChatGPT"), run across 4 engines (ChatGPT, Claude, Gemini, Perplexity), in 3 languages (English, Portuguese, Spanish). That's 20 × 4 × 3 = 240 total responses, the full denominator.

The numerator, per brand, counted as the number of those 240 responses that cited each name:

• Semrush: 80 responses cited it → 80 ÷ 240 = 33.3% AI SOV
• Profound: 60 responses → 25.0%
• Otterly.AI: 56 responses → 23.3%
• Peec AI: 47 responses → 19.6%
• Ahrefs: 47 responses → 19.6%
• SE Ranking: 37 responses → 15.4%
• Scrunch AI: 18 responses → 7.5%
• Writesonic: 12 responses → 5.0%
• Geoptie: 12 responses → 5.0%
• AthenaHQ: 8 responses → 3.3%
• LLMrefs: 7 responses → 2.9%
• Rankscale: 7 responses → 2.9%

Notice that these percentages don't sum to 100%. That's expected and correct: a single AI answer routinely cites multiple sources at once, so the same response can count toward several brands' numerators simultaneously. AI SOV, unlike a market-share pie chart, is not zero-sum by construction.

Per-Engine AI Share of Voice: Where the Real Story Lives

A single blended AI SOV number, like the leaderboard above, hides something the per-engine breakdown makes obvious: the same brand can have a wildly different score depending on which engine you're measuring.

Semrush's AI SOV by engine: ChatGPT 17%, Claude 35%, Gemini 38%, Perplexity 43%. That's a 26-point spread on the identical brand, identical category, engine as the only variable.

Otterly.AI's AI SOV by engine: ChatGPT 8%, Claude (not in the top-5 for that engine in our data), Gemini 32%, Perplexity 42%. Nearly a 5x difference between its weakest and strongest engine.

Ahrefs' AI SOV by engine: ChatGPT 13%, Claude 18%, Gemini (not top-5), Perplexity 28%.

The practical consequence: a blended AI SOV score of, say, 20% could mean "consistently around 20% everywhere" or "0% on ChatGPT and 40% on Perplexity," and those are two completely different competitive positions requiring two completely different strategies. Always calculate AI SOV per engine before drawing a conclusion from the blended number alone.

Per-Language AI Share of Voice: The Second Hidden Variable

Language shifts the number almost as much as engine does. In our data, Semrush's own AI SOV moved from 40% in English down to 35% in Portuguese and 25% in Spanish, a 15-point swing on the same brand, same category, language as the only variable. That's consistent with the models simply having different amounts of training exposure to a still-young category depending on language, not a real difference in the underlying product.

If a meaningful share of your buyers research in a language other than English, calculate AI SOV separately for that language rather than assuming your English-language number generalizes. A brand's English AI SOV and its Spanish AI SOV can tell two different competitive stories, and an English-only measurement will systematically miss whichever one applies to the buyers you're not measuring.

AI Share of Voice vs. Organic Footprint: They're Not the Same Game

One of the clearest patterns in our full dataset is that AI Share of Voice and organic search footprint are only loosely related, and sometimes not related at all. SE Ranking ranks organically for 27,823 keywords in our category scan, more than every other tool in the dataset combined, and its AI SOV was 15.4%, behind four smaller, GEO-native competitors with a fraction of its organic footprint. Profound, by contrast, ranks for 1,270 category keywords (roughly 22 times fewer than SE Ranking) and posted a 25.0% AI SOV, well ahead of SE Ranking despite the much smaller organic base.

The takeaway isn't that organic SEO doesn't matter, it's that AI Share of Voice is measuring something organic rank tracking doesn't capture: whether a model's synthesized answer chooses to name you, which depends on factors like topical authority, structured content, and training-data exposure that don't map cleanly onto keyword-by-keyword ranking position. Treat AI SOV as its own metric with its own tracking cadence, not a proxy you can infer from an organic rank tracker.

Detection Method Matters: Domain Citation vs. Name-Match

There are two ways to detect a "mention" in an AI response, and they carry different reliability:

• Cited source domain (the harder signal). The engine explicitly links to your domain as a source. This is the more reliable detection method because it's unambiguous, either your domain shows up in the citations or it doesn't.
• Name-match in the answer text (the softer signal). The brand's name appears somewhere in the generated prose, without necessarily linking to a source. This is useful for catching mentions that don't come with a citation, but it carries a real limitation: common-word brand names (think of a hypothetical tool literally named "Rank" or "Scale") can produce false positives when the word appears in an unrelated sentence.

Our own measurement combines both signals, but if you're building your own AI SOV tracking, be explicit about which signal (or combination) you're using, and disclose it when you report the number. A share-of-voice figure built purely on loose name-matching will run measurably higher, and less trustworthy, than one anchored primarily to cited-domain detection.

Setting a Realistic AI Share of Voice Target

Classic share of voice guidance often suggests targeting a SOV somewhat higher than your current market share, on the logic that share of voice today is a leading indicator of market share tomorrow. That logic transfers reasonably well to AI Share of Voice, with one adjustment: because AI SOV is measured against a specific, curated prompt set rather than the full, uncountable universe of channel impressions, your target should account for how many real competitors exist in your prompt set, not an abstract market-wide share.

A simple way to set a first target: take 100%, divide by the number of credible named competitors in your category (not counting long-tail players cited once or twice), and treat that as your "fair share" baseline. In our own category, with roughly 6 to 8 credible, frequently-cited competitors, a fair-share baseline lands somewhere around 12 to 17%, a more grounded near-term target than aiming straight for the category leader's share. Once you're consistently hitting fair share, a more ambitious target (matching or exceeding the category leader) becomes the reasonable next milestone, rather than the starting goal.

How to Calculate Your Own AI Share of Voice: Step by Step

1. Define your competitive set. List every brand a buyer would realistically consider alongside you, not just your two or three closest rivals. Our own set spans 17 named players, from category leaders to tools cited in a single response.
2. Build a real prompt list. Ten to twenty buying-intent prompts, phrased the way an actual buyer would ask, not generic category keywords. "What are the best [category] tools" is a fine starting prompt; "is there a free way to test [competitor] before switching" reflects a real, later-stage buying question.
3. Choose your engines and languages. At minimum, cover the engines with the highest research volume for your category. If a meaningful share of your buyers use a non-English language, run the same prompt set translated, don't assume your English result generalizes (see the per-language section above).
4. Run every prompt through every engine, in every language, and log the result. For each response, record which brands were cited, by domain, by name-match, or both, and keep the raw responses for auditing later.
5. Calculate the percentage. Citations for each brand, divided by total responses in your measurement set, times 100. Do this for the blended total, then again per engine and per language.
6. Re-run on a fixed cadence. Monthly is a reasonable default. A single run is a snapshot, not an average, treat it accordingly, and never present a one-time measurement as a stable trend line.

Tools for Measuring AI Share of Voice

• Manual, spreadsheet-based. Zero cost beyond your own time. Run each prompt yourself in each engine's chat interface, log results by hand. Workable for a small prompt set (10 to 20 prompts) measured monthly, but doesn't scale past that without becoming a part-time job.
• GEO-native tracking tools (Profound, Otterly.AI, Peec AI, and similarly-focused platforms). Built specifically around this measurement, typically with automated prompt scheduling, per-engine breakdowns, and citation-source detection built in as the core product rather than a bolt-on.
• Broader AI-visibility modules inside larger SEO suites (Semrush's AI Visibility Toolkit, similar add-ons from other incumbents). Convenient if you already pay for the rest of the suite, at the cost of typically less per-prompt and per-engine granularity than a purpose-built GEO tool.

Common Mistakes That Distort the Number

• Treating a single run as a stable score. AI model outputs vary run to run even on an identical prompt. A one-time measurement is a snapshot; report it as one.
• Using generic, low-effort prompts. A prompt set built entirely from category head terms inflates everyone's citation rate relative to the specific, comparison-heavy questions real buyers actually ask.
• Ignoring the engine breakdown. A blended AI SOV can mask a brand that's completely absent on the one engine your buyers actually use.
• Counting only branded queries. If every prompt already contains your competitor's name, you're measuring "does the engine repeat back the name I fed it" rather than "does the engine recommend this brand unprompted," a materially easier bar to clear.
• Skipping the language breakdown for a global brand. As shown above, the same brand's AI SOV can swing 15 points between languages. An English-only measurement quietly assumes every market behaves the same, and our data says that assumption is wrong.

Checklist: Your First AI Share of Voice Report

• [ ] Competitive set defined, 8 to 15 named brands, not just your top 2 or 3 rivals
• [ ] 10 to 20 buying-intent prompts written in real buyer language
• [ ] Engines chosen (minimum: the highest-research-volume engine for your category)
• [ ] Languages chosen (beyond English, if a meaningful buyer share isn't English-first)
• [ ] Detection method decided and disclosed (domain citation, name-match, or both)
• [ ] Blended AI SOV calculated
• [ ] Per-engine AI SOV calculated separately
• [ ] Per-language AI SOV calculated separately, if applicable
• [ ] Re-measurement cadence set (monthly is a reasonable default)
• [ ] Raw responses logged for audit, not just the final percentage

Related Reading

For the full leaderboard and per-engine breakdown behind every number in this piece, see which brands AI engines actually recommend. For the process of running a structured, repeatable citation check on your own site, see how to run a GEO audit and how to set up prompt monitoring. For a related metric that's easy to confuse with AI SOV, see our explainer on zero-click search and what it means for GEO.

---

The AI Share of Voice figures in this piece come from original research by the GeoHero Research Team: 240 AI-engine responses across ChatGPT, Claude, Gemini, and Perplexity (20 buying-intent prompts, three languages, July 2026), and a competitor organic-ranking scan of 14 domains in the category (our search-index scan, July 2026). Every percentage reflects a single measurement run, not an average across repeated runs. We re-run this monthly. Want your own AI Share of Voice baseline without building the spreadsheet yourself? Get the full report.