What Is a Good Share of Voice Percentage for AI Answers?
No single AI share of voice percentage is good for every brand. What changes the number, what published benchmarks rest on, and how to set a target from your own baseline.
Angelique Swain, AI-First SEO Strategist
· 6 min read
On this page
People ask what is a good share of voice percentage for AI answers from ChatGPT, Gemini and Google's AI Overviews, and the plain reply is that no single figure fits every brand. A good share beats your own baseline and your closest rivals' shares, such as Rival B's, on the same ChatGPT prompts, then holds across comparable runs. OptimizeGEO, HG Insights and Shadow publish thresholds, but they differ in what they divide by, and none shows the data behind it.
Is your 20% strong or weak? Against which rivals, on which prompts in ChatGPT or Gemini, on how many answers? And what would a target even be measured against in Firemark?
Answer those three and the number stops being a verdict and starts being a tool in Firemark or a Google Sheet.
Key takeaways
- No percentage is good on its own. 20% beside Rival B's 35% only means something on the same ChatGPT and Gemini prompts.
- Six things move the number without any change in the brand, from the prompt basket to a sample of 40 versus 200 answers.
- OptimizeGEO, HG Insights and Shadow use different divisors, and none discloses its prompts, sample or dates.
- Write a target as a gap, a base and a range, and grow the base from 40 answers toward 200 before you promise the gap.
What is a good share of voice percentage?
No single percentage is good for every brand. A share is good when it sits above your own baseline and above your rivals' shares on the same prompts in ChatGPT, Gemini and Google's AI Overviews, and when it holds across runs.
Firemark's method page says the same in one line: there's no universal good score, because the figure depends on your prompts, your market and your competitors. The definition and the arithmetic live in Firemark's guide to AI share of voice calculation, so this piece skips them.
It answers the question that Firemark's guide leaves open. If 30% can be strong in one market and weak in another, what do you compare yours with?
What moves the number?
Six things change a share of voice without any change in the brand: the prompt basket, the divisor, the number of rivals, the engines (ChatGPT, Gemini and Google AI Overviews), whether prompts name the brand, and the size of the sample. Each one is a reason to distrust a number quoted without its conditions.
| Prompt basket | Effect on the share: A basket of comparison questions and a basket of definition questions name different brands. Check before you compare: Same prompts on both sides |
|---|---|
| Divisor | Effect on the share: Answers collected as the base gives shares that can total over 100%. Total mentions as the base totals 100%. Check before you compare: Which one was used |
| Number of rivals | Effect on the share: More brands share the same answers. Check before you compare: How many brands are tracked |
| Engines | Effect on the share: Each engine answers differently, so a blend hides gaps. Check before you compare: Which engines, and the split |
| Branded prompts | Effect on the share: A prompt that contains your name makes a mention likely. Check before you compare: Whether they were counted |
| Sample size | Effect on the share: A small base gives a wide range. Check before you compare: The base and the range |
The rival count is plain arithmetic. If ChatGPT named every brand equally often and the divisor were total mentions, each of 4 brands would get 25%, each of 10 would get 10% and each of 20 would get 5%. Nobody lost a single answer, and the share still shrank because more brands arrived.
Branded prompts have a switch. Firemark can leave out the prompts whose wording names your brand, an alias or your website, on every screen, in the CSV and in the MCP, so you can read a category-only share beside the blended one.
Sample size matters most. Firemark's method page shows a 40% share with a 95% range of 23% to 59% on 25 answers and 36% to 44% on 500, and labels the table arithmetic, not a real brand.
Do published benchmarks help?
Published benchmarks don't help yet, because the guides from OptimizeGEO, HG Insights and Shadow that give percentages divide by different things and none shows the prompts, sample, dates or engines behind its thresholds.
Three vendor guides, read on November 5, 2026, give these ranges. They're rules of thumb, and each one says so in its own words.
| OptimizeGEO | What it calls good: 25% to 40% competitive, above 40% strong. What it divides by: Unclear: its formula says citations, its steps say responses naming the brand |
|---|---|
| HG Insights | What it calls good: 20% to 40% competitive, above 40% category leadership. What it divides by: Total mentions of every brand |
| Shadow | What it calls good: Leader 40% to 60% in a niche with 3 to 5 competitors, 20% to 35% with 10 to 20, 10% to 20% with 50 or more. What it divides by: Competitors' combined appearances in its worked example, so its shares don't total 100% |
Look at the third column: three divisors mean the percentages from OptimizeGEO, HG Insights and Shadow aren't interchangeable. Where the divisor is total mentions, shares add up to 100%, so an even split gives each brand 25% in a field of four and under 7% in a field of fifteen. HG Insights says as much, calling 20% unimpressive for the leader of a four-vendor category and remarkable for the seventh player in a field of fifteen.
Profound's own walkthrough shows how a large field compresses the numbers. In its example the top 10 brands' shares of voice vary by only about 2%, and the tracked brand ranks 344th. AthenaHQ advertises industry benchmarks in a report behind a sign-up form, so none of its figures appear here.
The one independent source is SparkToro's study with Gumshoe, an AI tracking vendor, a conflict the author states himself. It covered ChatGPT, Claude and Google's AI across 12 prompts and 2,961 runs, and isn't peer-reviewed.
Volunteers got the same list of brands twice less than 1 time in 100 on ChatGPT and Google's AI. SparkToro's author still called visibility percent, across dozens to hundreds of prompts run several times, a reasonable metric, and said more data is needed.
None of the thresholds from OptimizeGEO, HG Insights or Shadow comes with disclosed prompts, sample, dates and method, and no other public benchmark did either. Treat any of those thresholds as a conversation starter, not a target.
How do you set your own target?
Set a target against three things you can measure yourself: your baseline, a named set of rivals on the same prompts, and movement across comparable runs. A number you can't reproduce next month in ChatGPT or Gemini isn't a target.
- Fix the basket and the conditions: Keep the prompts, the engines and the location the same, and decide up front whether prompts that name your brand count. Firemark's guide to building a prompt set covers where the questions come from.
- Record the baseline over a base: Write "named in 8 of 40 answers", then add the 95% range. A bare percentage like Firemark's 40% hides how little it rests on.
- Read your rivals on the same prompts: Pick the nearest one or two, such as Rival B, as the comparison. Rival B's share on your basket is the one benchmark that shares your conditions.
- Write the target as a gap or a place: "Close half the gap to Rival B" or "move from third to second on middle-funnel prompts" can be checked. "Reach 40%" can't be defended.
- Judge it by the run rule: Firemark calls a movement only after three comparable runs and a change of 5 points or more, so a single good week doesn't count.
Advertising planners have an old version of this idea. Excess share of voice means buying a bigger share of the ads than your share of the market, in order to grow that market share, as an IPA blog post explains. That IPA research is built on ad spend, and nothing published links an AI share to market share, so use the IPA comparison as a sanity check on your ambition, not as a target.
What does a target look like?
A target is a gap, a base and a range, written together. The numbers below are invented to show the arithmetic.
Say your brand is named in 8 of 40 answers on one engine, say ChatGPT, and Rival B in 14. Those are shares of answers, 20% and 35%, so don't set them beside the OptimizeGEO, HG Insights or Shadow thresholds above. A 15 point gap looks like a clear goal.
| 40 answers | You: 8 named, 20%, range 10% to 35%. Rival B: 14 named, 35%, range 22% to 50% |
|---|---|
| 200 answers, same shares | You: 40 named, 20%, range 15% to 26%. Rival B: 70 named, 35%, range 29% to 42% |
On 40 answers the two ranges overlap, so the 15 point gap to Rival B could be noise, and a target to close it can't be called either way. On 200 answers the ranges separate, so the same 15 points sits outside sampling noise, assuming the answers are independent.
Repeat runs of the same ChatGPT prompts narrow the range less than new prompts do. The ranges use the Wilson method for confidence intervals for proportions, which the NIST handbook describes.
So grow the base from 40 to 200 before you promise the gap. Report the target as "named in 55 or more of 200 answers, up from 40, with Rival B at 70", not as a percentage alone.
Frequently asked questions
Is 20% a good AI share of voice?
A 20% share can be strong or weak, depending on the divisor, the number of rivals and the prompts. With total mentions as the divisor, an even split gives each of 4 brands 25%, so 20% is below it. In a field of fifteen, an even split gives 6.7%, so 20% is far above it. Compare it with Rival B's share on the same ChatGPT and Gemini prompts.
Is a higher share always better?
A higher share isn't always better, because it depends on what the prompts contain. Prompts that name your brand raise the count without telling you whether buyers who don't know you hear your name in ChatGPT or Gemini. Read the share by funnel stage in Firemark, and keep named and linked answers apart.
How many prompts do I need before I set a target?
You need enough answers that your range separates you from the rival you're comparing with. Firemark's method page shows a 40% share spanning 23% to 59% on 25 answers and 31% to 50% on 100. Start with the base that makes the gap you care about readable.
How often should I re-check the target?
Re-check the target after each comparable run, and judge movement over several. Firemark calls a change only after three comparable runs and a move of 5 points or more. A change in how ChatGPT or Gemini answers are collected restarts the comparison. Firemark's five-step measurement method covers the first baseline.
Can Search Console show my AI share of voice?
Search Console can't show a share of voice. Its Generative AI performance report counts impressions for links to your site in AI Overviews and AI Mode, with no rivals and no record of whether the answer named your brand.
Find out where your share stands
A good share of voice is the one you can reproduce, compare with a named rival and defend with a base and a range. Firemark counts named and linked answers over a base for each brand you track, draws a 95% range per engine and measures what ChatGPT, Gemini and Google's AI Overviews say without changing what any engine shows. Open Firemark's demo report to see a run on sample data, or start with a small set of prompts and a rival or two.
