How many prompts do I need?

Twenty to forty is a practical range. Enough to cover your main buyer questions and smooth out randomness, few enough to run regularly. Consistency of the set matters more than its size.

Should I weight mentions by position?

You can. A top recommendation is worth more than a name buried in a list. Start with a simple presence count for clarity, then add weighting once the basic trend is stable.

The same prompt gives different answers each time. How do I handle that?

Run each prompt a few times and record what appears across runs. Some variation is normal. Patterns across the whole set, not any single answer, are the signal.

Yes. The steps are mechanical, which makes them a good fit for software that runs the prompts, logs the brands, and tracks the trend for you.

How to Measure Share of Voice in AI Answers

You can feel that AI answers matter without knowing where you stand in them. Share of voice turns that hunch into a number: how often the engines name you, measured against how often they name your rivals. This guide shows how to calculate it and keep it current.

What is share of voice in AI answers?

Share of voice is your slice of the mentions. Take the questions your buyers ask, run them through the engines, and count how often each brand shows up. Your share of voice is your mentions divided by the total mentions across you and your competitors. A rising share means the engines are naming you more. A falling one is an early warning.

How do you calculate it?

The method is simple and repeatable. The value comes from running it the same way every time.

Fix your prompt set. Write twenty to forty questions a buyer would ask, and keep the list stable so you can compare month to month. Include category questions, alternatives-to-a-rival questions, and direct is-this-right-for-me questions.
Choose your engines. Start with ChatGPT, Claude, and Google's AI answers. Run each prompt on each, since the same question returns different names on different engines.
Log every brand named. For each answer, record which brands appear, including yours. Note the order and any description, but the core count is presence.
Do the math. Suppose you run 30 prompts on one engine. Your brand appears in 12 answers and your main rival in 21. Against that rival, your share is 12 out of 33 combined mentions, roughly a third. Repeat per rival and per engine.
Record the baseline. Save the numbers with the date. One reading is a snapshot; the trend is what you act on.

What counts as a mention?

Count a mention when the engine names your brand in a relevant answer. Decide the edge cases up front and hold to them. A passing reference in a long list still counts, though it is weaker than a top recommendation. Track a wrong description as a mention with a flag, because being named incorrectly is a problem to fix, not a win to log.

How often should you measure?

Monthly is enough to see real movement without chasing noise. Re-run the same prompts, compare to your baseline, and look for two things: brands entering or leaving the answers, and your own description drifting. Ship a major change or watch a competitor launch, and run an extra check. For the tactics that move the number, see how to get your brand mentioned in ChatGPT answers.

How do you turn the number into action?

A share-of-voice figure is only useful if it points somewhere. Read it against two things: your rivals and your own history.

Against rivals, a low share tells you the engines lean on sources that favor them. Find the answers where a competitor appears and you do not, then look at what those answers cite. It usually traces to a review site, a comparison page, or a piece of coverage you are missing. That gap is your work list.

Against your own history, a falling share is an early warning even when the absolute number still looks fine. A brand that drops from a third of mentions to a fifth over two months is losing ground before any traffic report would show it. Treat a downward trend as a prompt to check what changed: a competitor's new content, a refreshed model, or a fact of yours that went stale.

Prioritize by buyer value. A slip on a high-intent question, the kind a ready buyer asks, matters more than a slip on a broad one. Fix the answers closest to the sale first.

How do you sharpen the score over time?

Add weight once the basic count is stable. A first-position recommendation is worth more than a name dropped at the end of a list, so score a top mention as a full point and a trailing one as a half. Track position too, since sliding from first-named to last-named inside an answer is a real loss even when your presence count holds. Read the citations under each answer where the engine shows them, because the sources it leans on are your shortest path to changing what it says.

Set a review trigger you will actually honor. A sharp drop in share against a key rival, or falling out of three or more answers you used to appear in, is worth a same-week look rather than waiting for the monthly pass. Write the trigger down, so a bad month prompts action instead of a shrug.

The measurement loop

Freeze a set of 20 to 40 buyer questions.
Run them across ChatGPT, Claude, and Google.
Log which brands each answer names.
Divide your mentions by the total to get share of voice.
Re-run monthly and watch the trend, not the snapshot.

Doing this by hand across several engines and dozens of prompts is a real chore, which is why most teams do it once and never again. Programmatic CMO runs the set on a schedule through its GEO agent and charts the trend, so the number stays live. For the concept behind it, start with what generative engine optimization is.

How to measure share of voice in AI answers