How do you calculate share of voice in AI search answers?
Share of voice in AI search is the number of answers that name a brand divided by the same count summed over every brand you track, in the same answers: SOV = answers naming the brand / Σ answers naming each tracked brand. Report it next to the mention rate and the citation rate, because the citation rate can put brands in a different order: in our monitoring of ChatGPT answers to web search API questions, SerpApi had 8.2% share of voice and was named in 33.5% of answers but cited in only 7.8%, while cloro had 0.3% share of voice and was cited in 37.6%.
- Within one engine, share of voice and mention rate give the same order. Share of voice is the mention rate divided by a number that is the same for every brand in that slice. The two differ when you compare engines or periods.
- Across engines they can move in opposite directions. Tavily was named in more Bing AI Mode answers than Google AI Mode answers (80.8% against 79.6%), but its share of voice was lower on Bing AI Mode (16.9% against 23.3%), because Bing AI Mode answers named more brands each.
- Citation rate is a different ranking. On three of five engines, the brand with the highest citation rate was not the brand with the highest share of voice among the five compared.
To calculate share of voice in AI search answers:
- Fix the prompt set, the engines and the panel: the list of brands that make up the denominator. Keep all three fixed for the whole period you report.
- Run every prompt on every engine on a schedule and store each answer with its text and its cited links.
- For each answer and each panel brand, record two flags: named (the brand appears in the answer text, with link text removed) and cited (a brand domain is among the cited links).
- Count the answers that name each brand. Share of voice is that count divided by the sum of the counts over the panel.
- With the same answers as the denominator, compute the mention rate (answers naming the brand / answers) and the citation rate (answers citing the brand / answers).
- Report the three per engine, with the panel listed, and restate the history when the panel changes.
How do you calculate share of voice in AI search answers?
Three measures come from the same set of answers:
| Measure | Formula | Question it answers |
|---|---|---|
| Share of voice | answers naming the brand / Σ answers naming each panel brand | Of all the brand mentions in these answers, what part is ours? |
| Mention rate | answers naming the brand / answers | How often does an answer name us? |
| Citation rate | answers citing a brand domain / answers | How often does an answer link to us as a source? |
Each answer counts once per brand, so a brand named five times in one answer counts the same as a brand named once. A count-weighted variant sums every occurrence instead, and a position-weighted variant gives more weight to brands named first. Both reward long answers that repeat a name, so state which one you use.
Share of voice depends on the panel. Add a competitor and every other brand's share goes down, even if no answer changed. Mention rate and citation rate do not depend on the panel.
The query below computes all three from the tables of the open-source geo-tracker, which stores one row per answer and per brand, including the brands that were not named:
WITH answers AS ( -- results that carry an answer: AI prose or an AI Overview SELECT r.id FROM results r WHERE r.status = 'completed' AND r.completed_at > now() - interval '30 days' AND r.engine IN ('chatgpt', 'gemini', 'aimode', 'bing-aimode', 'google') AND coalesce(r.response #>> '{result,text}', r.response #>> '{result,aioverview,text}', '') <> '' ), scored AS ( SELECT b.name, m.mentioned, EXISTS ( -- a brand domain among the answer's citations SELECT 1 FROM result_sources s, unnest(b.domains) AS d(domain) WHERE s.result_id = m.result_id AND s.kind IN ('source', 'citation_pill', 'ai_overview') AND (s.domain = lower(d.domain) OR s.domain LIKE '%.' || lower(d.domain)) ) AS cited FROM result_brand_mentions m JOIN answers a ON a.id = m.result_id JOIN brands b ON b.id = m.brand_id WHERE b.enabled ), per_brand AS ( SELECT name, count(*) AS answers, count(*) FILTER (WHERE mentioned) AS named, count(*) FILTER (WHERE cited) AS cited FROM scored GROUP BY name ) SELECT name AS brand, round(100.0 * named / answers, 1) AS mention_rate_pct, round(100.0 * cited / answers, 1) AS citation_rate_pct, round(100.0 * named / nullif(sum(named) OVER (), 0), 1) AS share_of_voice_pct FROM per_brand ORDER BY share_of_voice_pct DESC NULLS LAST; The denominator leaves out Google results with no AI Overview, because a brand cannot be named in an answer that was not shown. The citation check reads only the links the answer cites, not organic results or ads on the same page.
Why do share of voice, mention rate and citation rate rank brands differently?
Because a brand can be named without a link, and linked without being named. We ran a fixed set of web search API questions once a day on five engines from October 17 to November 15, 2026, with a panel of ten vendors. On ChatGPT:
| Brand | Share of voice | Mention rate | Citation rate | Rank by SOV | Rank by citation |
|---|---|---|---|---|---|
| Brave Search | 20.3% | 82.9% | 42.4% | 1 | 1 |
| Tavily | 19.6% | 80.0% | 33.5% | 2 | 4 |
| Exa | 19.2% | 78.4% | 38.8% | 3 | 2 |
| SerpApi | 8.2% | 33.5% | 7.8% | 4 | 5 |
| cloro | 0.3% | 1.2% | 37.6% | 5 | 3 |
Ranks are among the five brands shown. The other five panel brands (Firecrawl, Parallel, Serper, Perplexity's Sonar API and Linkup) make up the rest of the share of voice.
Share of voice and mention rate put the five in the same order, as they must inside one engine. The citation rate does not. SerpApi is named in a third of the answers and cited in fewer than one in twelve: engines know the name but rarely link its pages for these questions. cloro is the opposite case, cited in more than a third of the answers and named in almost none, so the engine cites its pages and rarely names it. A share-of-voice report alone would miss both.
The same split shows on every engine. Share of voice, then citation rate, for the same five brands:
| Engine | Brave Search | Tavily | Exa | SerpApi | cloro |
|---|---|---|---|---|---|
| Google AI Overview | 17.3% / 16.0% | 20.4% / 11.5% | 19.1% / 18.9% | 8.9% / 9.0% | 0.3% / 10.2% |
| Google AI Mode | 20.2% / 23.2% | 23.3% / 27.6% | 19.6% / 18.4% | 8.0% / 4.8% | 0.2% / 3.2% |
| ChatGPT | 20.3% / 42.4% | 19.6% / 33.5% | 19.2% / 38.8% | 8.2% / 7.8% | 0.3% / 37.6% |
| Gemini | 18.5% / 18.0% | 18.8% / 12.4% | 18.6% / 17.6% | 8.9% / 0.4% | 0.1% / 4.4% |
| Bing AI Mode | 13.8% / 18.3% | 16.9% / 15.8% | 15.9% / 32.5% | 10.6% / 1.7% | 1.0% / 24.2% |
Tavily leads share of voice on four engines and leads citation rate on one. Exa leads citation rate on two. On Gemini, SerpApi keeps 8.9% of the voice with almost no citations.
How does the length of an answer change share of voice?
Share of voice divides by the total number of panel brands named, so it falls when answers list more brands. The mention rate divided by the share of voice gives the average number of panel brands per answer. For Tavily, that is 3.4 on Google AI Mode (79.6% / 23.3%) and 4.8 on Bing AI Mode (80.8% / 16.9%). Tavily is named in a slightly larger share of Bing AI Mode answers, but each of those answers names more competitors, so its share of voice is lower.
This is the reason to report the mention rate next to share of voice. A drop in share of voice can mean that the brand is named less, or that answers got longer, or that a competitor was added to the panel. Only the mention rate separates the first case from the other two.
How many answers does a share of voice number need?
As many as a mention rate does, because it is built from one. Each answer either names the brand or not, so a rate near 30% needs about 81 answers for a 95% interval of ±10 points and about 323 for ±5 points (n = 1.96² × p × (1 - p) / E²). Answers for the same prompt also change from day to day, so spread the runs over days and prompts rather than repeating one prompt in one hour. A brand at 0.3% share of voice, like cloro on ChatGPT above, is within noise of zero on any sample a team would collect in a week.
How was this checked?
- Data. Our own monitoring of the prompts in one topic, search APIs for AI agents and RAG, one run per prompt, engine and day, October 17 to November 15, 2026, through logged-out sessions. Bing AI Mode joined on November 7, so its row covers 9 days.
- Panel. Ten vendors, fixed for the whole window: cloro, Tavily, Exa, Parallel, Brave Search, SerpApi, Perplexity's Sonar API, Firecrawl, Serper and Linkup.
- Named. A deterministic, word-bounded match on the answer text with link text and bare citation labels removed, so a link does not count as a mention.
- Cited. A source URL whose hostname ends in the brand's domain (brave.com, tavily.com, exa.ai, serpapi.com, cloro.dev). For Brave Search, brave.com also covers pages about the Brave browser.
- Answers. Results with a shown answer. For Google AI Overview, results where an AI Overview was shown.
This post is published by cloro, which sells an API that returns AI engine answers as structured data, and is one of the five brands in the tables. The cloro rows come from the same query as the other rows. More on the formulas and their variants is in the AI share of voice guide.
原文:https://www.a8q9.cn/en/article/how-do-you-calculate-share-of-voice-in-ai-search-answers-f9e619.html