Share of voice in AI answers: what it is and how to measure it

Almost every SaaS founder has done it: opened ChatGPT, typed "what's the best tool for…", and felt either relief or panic depending on the answer. The problem is that the next answer can be different. So can the answer in Gemini, or the one your buyer sees tomorrow.

Share of voice is how you turn that anecdote into a number.

Why a single check is misleading

AI answers are not fixed lists. Three things make one-off checks unreliable:

• Answers vary between runs. Ask the same question three times and the brands, their order and the wording can all change.
• Models disagree. ChatGPT, Gemini, Claude, Perplexity, Grok and Google's AI Overview draw on different sources and recommend different brands.
• Wording matters. "Best tool for X" and "alternatives to Y" are different intents, and they surface different products.

A useful measurement has to account for all three.

Step 1: Build the question set

Start from what your buyers actually ask, not from your keywords. Cover four kinds of intent:

• Category: "What's the best competitor tracking tool for a SaaS team?"
• Problem: "How do I know when a competitor changes their pricing?"
• Comparison: "Vantix vs Retailo for a small team?"
• Alternatives: "What are good alternatives to Vantix?"

Thirty to a hundred questions is a good range. Fewer and one odd answer skews the result; many more and you are measuring noise in your own question list. Keep the set fixed, so this week's number is comparable with last month's.

Step 2: Ask every surface, more than once

Run every question on each AI surface your buyers use, and sample each one three times. Treat the result as a frequency — named in two of three runs is 0.67, not "yes".

Step 3: Extract mentions and positions

For every answer, record which brands were named and in what order. Do it with deterministic matching (your brand name and its known variants), not by asking another model to judge. A judge model adds its own variance on top of the variance you are trying to measure.

Step 4: Calculate the score

A simple, transparent formula works well:

• Mention rate = answers that name you ÷ all answers
• Position weight = 1 for first place, 0.5 for second, 0.33 for third, and so on
• Share of voice = your weighted mentions ÷ everyone's weighted mentions

Because the weights are open, anyone on your team can check how a number was reached. The same input always gives the same score.

What is a good number?

There is no universal benchmark, and anyone who quotes one is guessing. Two comparisons matter instead:

• Against your competitors. 20% is excellent in a crowded category with ten credible tools and weak in a category with three.
• Against yourself over time. The trend tells you whether your work is landing. A score that rises after you publish a comparison page, or appear in a review roundup, is the clearest feedback loop you will get.

Common mistakes

• Counting citations as mentions. Being linked as a source and being recommended are different things. Track them separately.
• Ambiguous brand names. If your name is a common word, matching needs care, or you will count sentences that aren't about you.
• Changing the question set. Every edit breaks comparability with past weeks. Add questions in batches and keep a note of when.

Scooutly runs this whole method on a schedule: a question set built from your category, six AI surfaces, three samples per question, deterministic brand matching and an open scoring formula. You see your share of voice next to your competitors', which models recommend you, and which sources the answers lean on — so you know what to change, not just how you scored.