How to measure AI share of voice

To measure AI share of voice, run a fixed basket of prompts against the AI assistants your buyers use (ChatGPT™, Claude®, Gemini™, Perplexity™), record which brands each answer mentions, and divide each brand's mention count by the total number of valid responses. If your brand appears in 18 of 90 responses, your mention rate is 20%; a competitor appearing in 45 of 90 holds 50%. Re-run the identical basket on a schedule and those rates become a trend you can manage against.

What counts as AI share of voice?

AI share of voice (SoV) is answer-level presence across a fixed prompt set: of all the responses the assistants generated, what fraction mention your brand, and what fraction mention each competitor. Unlike media share of voice, the denominator is not impressions or spend. It is the set of answers actually produced for your prompts.

Three variants matter in practice:

• Mention rate (the base metric). Responses that mention the brand, divided by total valid responses. Every brand is scored against the same denominator, so the rates are directly comparable.
• Position-weighted share. In list-style answers, being named first is worth more than being named ninth. A position-weighted variant gives full credit for top placement and partial credit further down the list.
• Sentiment-weighted share. A recommendation is worth more than a name in a comma-separated list. Weighting each mention by its tone turns raw presence into something closer to preference.

How do I measure it step by step?

1. Define the prompt basket. 20 to 40 questions your buyers actually ask, weighted heavily toward organic prompts (your brand name absent) with a smaller branded set.
2. Choose the providers. The assistants your audience uses; ChatGPT™, Claude®, Gemini™, and Perplexity™ are the common four.
3. Run every prompt against every provider at least 3 times. Identical prompts produce different answers run to run, so single runs are noisy.
4. Extract mentions. For each response, record which brands are named, in what position, and in what tone. This is the labor-intensive step and the one most worth automating.
5. Compute the rates. Mentions divided by valid responses, per brand, overall and per provider.
6. Repeat on a schedule with the identical basket. The trend is the product; a single snapshot is only a baseline.

The formula, with a worked example

The base formula in plain terms: a brand's AI share of voice equals the number of valid responses that mention the brand, divided by the total number of valid responses in the measurement cycle. Some teams also normalize to share of total mentions (your mentions divided by all tracked brands' mentions combined); that version always sums to 100% across brands, but the per-brand mention rate is easier to interpret and harder to distort, so treat it as primary.

Say you track 30 prompts and run the basket 3 times, for 90 total responses. Extraction finds your brand in 18 responses and Competitor A in 45.

Brand	Responses mentioning it	Mention rate
Your brand	18 of 90	20%
Competitor A	45 of 90	50%

Competitor A owns 2.5x your raw presence. Now factor in tone. Suppose 10 of your 18 mentions are neutral listings (your name in a list, no evaluation) and only 8 are actual recommendations, while most of Competitor A's 45 mentions are enthusiastic endorsements. The effective gap is then wider than the raw one, because a name in a comma-separated list does not steer a buyer the way an explicit recommendation does. That is what the sentiment-weighted variant captures, and it is usually the more honest picture of where buyers are being steered. At minimum, record the tone mix behind each brand's count rather than the count alone, and expect any tool you use to show you that mix.

What makes the number comparable over time?

Four method decisions, each easy to get wrong.

Fix the prompt basket

Never change prompts mid-series. Swap even a few and the trend line silently becomes a comparison between two different questionnaires. When the business needs new prompts, start a new tracked series or anchor comparisons to the shared subset rather than editing the old basket. We cover basket management in how to track the prompts that surface your brand.

Hold the provider set constant

Each assistant has its own retrieval behavior, source preferences, and answer style, so the same basket yields different rates per provider. Report per-provider rates alongside the blend, and never add or drop a provider mid-series: doing so moves every brand's number for reasons that have nothing to do with the market.

Run the basket more than once

Identical prompts produce different answers on different runs. A brand that appears in two of three runs looks absent in a one-run snapshot, so a single run can misstate its rate through sampling noise alone. Three runs per prompt per provider is a practical floor; average across runs. In our own multi-scan corpus, a single run surfaced just 78.5% of the brands three runs surfaced. The mechanics are in why AI answers change between runs.

Separate branded from organic prompts

A branded prompt contains your name ("Is Acme good for enterprise teams?"), so the answer almost always mentions you. That is useful for measuring knowledge depth, but blended into share of voice it structurally inflates your number, and no competitor comparison survives it. Report organic SoV as the competitive metric and keep branded results in their own column.

Why citation share is a second SoV lens

Prose mentions are one scarce resource in an AI answer; citation slots are another. Across a sample of scans we ran between April and August 2026 with web search enabled on every request (33,000+ responses from the four major AI providers), answers drew on more than 26,000 unique domains, and a typical single scan's answers cited a median of 237 distinct domains. No shortlist dominates that pool: the ten most-cited domains combined account for under 9% of all citations. When answers are grounded like this, each one is a vote on sources, and the electorate is enormous.

So track citation share alongside mention share: of all domains cited across the basket, what fraction belong to you versus your competitors. The two lenses disagree often enough to be worth separating. A brand can be mentioned without being cited (the model knows it from training data), and a domain can be cited without the brand being named prominently in the prose.

What tools help benchmark brand visibility across AI assistants?

You can prototype the method manually: a spreadsheet, 20 to 30 prompts, an afternoon of copy and paste, and a tally of mentions. That is worth doing once to build intuition. It does not survive contact with the real workload: 30 prompts times 3 runs times 4 providers is 360 responses per measurement cycle, each needing mention extraction, sentiment classification, and citation parsing.

That workload is what the category tools automate: tracked prompt sets executed across the major assistants, mention and citation extraction, and competitive share computed per provider. Tools differ in provider coverage, run counts, weighting schemes, and how much prompt-level detail they expose. Gen3 AI Visibility computes each brand's share from repeated runs per prompt and shows the tone mix and per-prompt detail behind every number. Whichever tool you evaluate, apply the comparability rules above: the dashboard number is only as trustworthy as the fixed basket, constant provider set, and multi-run sampling behind it.

Measure it

The fastest way to get a baseline is to see how the major assistants describe your brand today: run a free Pulse visibility check from the homepage (/). Then use the rest of the guides at /learn/ to turn that snapshot into a tracked share-of-voice series.