AI Share of Voice: How We Measure Brand Citations in AI Answers

Quick answer: AI share of voice is the percentage of sampled AI answers that name your brand. We put fixed buyer prompts to real engines with live web search, record mentions and returned sources, and repeat the runs. We report each engine separately, with a confidence range instead of a rank.

Ask ChatGPT or Google’s AI the same buyer question twice, and the brand list almost never matches. In a study by SparkToro and Gumshoe.ai, 600 volunteers ran 12 prompts a combined 2,961 times across ChatGPT, Claude and Google’s AI. For ChatGPT and Google’s AI, there was less than a 1 in 100 chance that two runs returned the same brand list.

That non-deterministic drift is the first of three measurement problems. The second is attribution: an answer can name your brand without citing your site, or cite your site without naming you. The third is the denominator: share of answers, share of tracked brand mentions and own-site citation counts are different numbers that often carry the same label. We measure mentions and citations separately, because one blended percentage hides what visibility means.

Here, AI share of voice means the answers that name your brand divided by all answers to a fixed prompt set, calculated separately for each engine. Competitive share of voice is your brand mentions divided by all tracked brand mentions.

Traditional share of voice uses a different denominator. Nielsen defines share of voice as a brand’s media spending as a percentage of all spending in its category, within a specific market, channel and time. We measure observed answers, not ad spend or search engine market share.

1. Build a fixed prompt set

We write a fixed set of buyer prompts for each brand, drawn from the brand’s own facts. Because the set stays the same, week-to-week numbers stay comparable. If the mix of questions changes, the measurement moves before any answer does.

By default, the headline number leaves out any prompt that contains the brand name. Ask an engine about a named company, and that company will show up. A buyer comparing options with no vendor in mind is a different situation.

1. We write buyer prompts from the brand’s facts.
2. We keep the prompt set fixed so results compare over time.
3. We leave brand-name prompts out of the headline percentage by default.

Five measures separate brand presence, source attribution, competitive coverage and the searches behind an answer. How often you check each one follows our engine sampling cadence. It is not continuous monitoring.

2. Sample Live, Search-Grounded Answer Engines

We send each prompt to real AI engines with live web search turned on, worded the way a person would type it. Each answer draws on fresh search results, a form of retrieval-augmented generation (RAG), rather than the model’s training data alone. Today we sample two engines this way: Gemini grounded in Google Search, and Claude with web search.

We add an engine only when an official interface returns its sources. We do not scrape consumer chat apps. Engines without that kind of interface, including Microsoft Copilot, sit outside the sample.

That boundary shapes how you read the result. A percentage from our Gemini sample describes those Gemini answers. It does not stand for every AI service, interface or buyer conversation.

First-party reporting answers a related question. Microsoft launched AI Performance in Bing Webmaster Tools as a public preview in February 2026. It shows a site’s citations across Microsoft Copilot, AI-generated summaries in Bing and select partner integrations.

Its metrics include total citations, average cited pages, grounding queries, page-level citations and visibility trends over time. Those numbers describe your own site’s citations. They do not show your competitive share across a shared set of buyer prompts.

3. Separate Unlinked Brand Mentions from Grounded Domain Citations

A mention is the brand named in the answer text, linked or not, and known aliases count. Exclusion phrases keep a common word that matches the brand name from being counted as a mention.

A citation is a separate record: the brand’s own domain appears in the sources the engine returned. We log the tracked competitors each answer names and every cited domain. For Gemini, we also capture the background searches it ran, known as query fan-out.

These rules keep each number precise. If an answer names the brand but its domain is missing from the sources, that counts toward mention rate only. If the brand’s domain is a returned source, that counts toward citation rate, even when the answer never names the brand.

We never ask a model to imagine an answer or report its own citations. We read the answer text and the engine’s source list. Our guide to tracking your brand in AI answers covers the questions teams need to keep apart when they monitor visibility.

4. Repeat the runs

Answers change between runs, so one answer is one observation, not a verdict on visibility. We ask each prompt again about every two days on Gemini and about once a week on Claude.

In the SparkToro and Gumshoe.ai study, Claude repeated brand lists slightly more often than ChatGPT and Google’s AI, but still rarely. The study recommends tracking visibility percentages across repeated runs instead of ranking positions. That is why we count whether a brand is present, not where it appears in the answer. The order shifts from run to run.

We read change prompt by prompt, comparing the last 7 days with the 30 days before. That ties every shift to one question: which buyer prompt produced different brand coverage? An overall percentage can’t answer that on its own.

5. Report a share with a range

Our headline number is the share of answers that name the brand, shown with a 95 percent range from a Wilson score interval. We never present it as a position or a rank.

The NIST handbook on confidence intervals for proportions describes the Wilson method and cites the statistical literature that recommends it. The simpler normal-approximation formula can produce impossible limits, such as a lower bound below zero.

For AI share of voice, the range puts sampling uncertainty in plain view. We treat any estimate built on fewer than about 400 answers as rough. A bigger sample narrows the range, but it does not make the prompt set stand for the whole market.

A useful report shows context next to the percentage: the engine, the prompt set, the answer count and the number of repeat runs. Ask how many times each prompt ran and how many distinct prompts were included.

Also ask whether the prompt set is disclosed. If you can’t see the questions, you can’t judge whether the percentage reflects how your buyers decide.

A hypothetical example with round numbers

This example is hypothetical. It is not a customer result or a benchmark.

Picture 400 sampled answers from one engine, with brand-name prompts excluded. Your brand appears in 200 answers, and your domain is a returned source in 80.

Mention rate is 200 divided by 400, or 50 percent. Citation rate is 80 divided by 400, or 20 percent. The 95 percent Wilson interval around the mention rate runs from about 45 to 55 percent.

Now say those same answers contain 800 tracked brand mentions in total, counting your brand and tracked competitors. Your competitive share is 200 divided by 800, or 25 percent.

Each percentage uses a different denominator and answers a different question. None of them proves a first-place position, and none measures search engine usage or market share.

Limitations of AI Share of Voice

A fixed prompt set makes results comparable, but it also limits what the measurement covers. Any question outside the set is outside the result. Repeat runs add observations, but they do not prove the chosen questions cover every buyer.

A confidence interval describes uncertainty around the measured share. It does not validate the choice of questions.

Engine coverage is a second boundary. Our sample covers Gemini grounded in Google Search and Claude with web search, so it says nothing about Microsoft Copilot. A domain in the source list and a brand named in the answer are separate observations, and they say different things about attribution.

Traffic attribution is a third boundary. Some AI apps send no referrer when a reader clicks a cited link, so analytics tools such as GA4 can file the visit under Direct. A reader who sees your brand in an answer and later types your domain also lands as Direct. A citation rate shows the engine returned your domain; it does not show how much pipeline followed.

Location is a fourth. Search-grounded answers draw on live search results, and those can differ by location and language. A sample run from one setup describes that setup, not every region your buyers search from.

Then there is the work itself. Collecting several hundred answers per engine and repeating them on a schedule means automated API calls, usage costs and rate limits. A spreadsheet and a few manual chats can’t keep that pace.

First-party search reporting does not close the competitive gap. Google Search Central says AI Overviews and AI Mode appearances are reported in Search Console’s Performance report, within the Web search type, alongside regular results. That shows your own search performance, not cross-brand coverage on a shared prompt set.

Use these measures side by side. Keep each one’s limits clear when you read it.

Common measurement mistakes

The worst mistakes change what a percentage means while the label stays the same. A clean chart still needs a clear denominator and real answers behind it.

• Mixing denominators: Share of answers and share of tracked brand mentions are different calculations. Label each one clearly.
• Including branded prompts: Questions that already name the company inflate discovery visibility. Report them separately.
• Treating one run as a ranking: The repeat-run evidence supports measuring presence across answers, not assigning a fixed position.
• Counting generic name matches: A matching word is not a mention when the answer means something other than the brand.
• Accepting unsupported citations: Check the sources the engine returned. Never ask a model to describe its own references.
• Comparing unlike reports: Own-site citation counts and cross-brand mention shares answer different business questions.

When you evaluate AI visibility analysis tools, start with these distinctions. Ask for the prompt set, the repeat-run counts and the uncertainty range before you treat a percentage as evidence for a decision.

What to Demand From a GEO Measurement Framework Before You Trust It

Before an AI share of voice number shapes a budget or a content plan, ask whoever produced it for five things: the full prompt set, the answer count per engine, the repeat runs per prompt, the returned source lists and the confidence range. If any of the five is missing, treat the percentage as a sketch and keep it out of your AEO KPIs until the gap is filled.

AEO Agent X tracks positions, clicks, indexation, leads and AI citations per URL. That product measure answers a different question from the brand-level method described here: which of your own pages engines cite, and how each one performs. Use the URL view for your pages and the brand-level share for competitive coverage.

Want a quick baseline first? Our free instant site report puts five buyer questions to an AI answer engine and shows whether your site was a source. When you’re ready, request access to AEO Agent X for approval-first marketing across SEO, AEO and GEO.

Frequently Asked Questions (FAQs)

How is AI share of voice calculated?

AI share of voice is the number of answers that name your brand divided by all answers to a fixed prompt set, calculated per engine. Competitive share is your brand mentions divided by all tracked brand mentions. Neither one measures search engine market share.

What is the difference between a brand mention and a citation in AI search?

A brand mention means the answer text names your brand, linked or not. A citation means the engine returned your domain as a source for the answer. An answer can name you without citing your site, or cite your page without naming you, so we report mention rate and citation rate as separate numbers.

How many AI answers make the measurement reliable?

We treat any estimate built on fewer than about 400 answers as rough, and we report a 95 percent Wilson interval with every number. A larger count still does not prove the prompts represent every buyer. Check the number of distinct prompts and the repeats per prompt alongside the total.

Does Google Search Console show AI Overview citations?

Yes, as part of your existing data. Google reports AI Overviews and AI Mode appearances in Search Console’s Performance report under the Web search type, alongside regular results. It does not give you a cross-brand citation share against competitors on a shared prompt set.

Does Bing Webmaster Tools show Microsoft Copilot citations?

Yes. The AI Performance public preview in Bing Webmaster Tools reports your site’s citations across Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. Its metrics include total citations, average cited pages, grounding queries and visibility trends. It does not show your competitive share across shared buyer prompts.