- Mention rate, citation rate, citation share and share of voice answer different questions. Each needs a stated numerator, denominator and prompt set.
- AI answers are not deterministic, so run every prompt several times and report the number of answers behind each figure.
- Small samples give wide error bars. Treat movement of a few points as noise unless it repeats across runs, engines and months.
Why the definitions matter
AI visibility metrics are ratios computed from a saved set of AI answers. Mention rate counts answers that name your brand, citation rate counts answers that link to your pages, citation share compares your links with everyone else's links, and share of voice compares your brand mentions with those of named competitors. Each is only meaningful when you state the prompt set, the engines, the dates and the number of answers used.
The same words are used in different ways across reports and tools. One report says share of voice and means the share of answers that mention you. Another means your mentions divided by all brand mentions. Both are legitimate, but they are different numbers. If you compare your figure with one from another source without checking the formula, you may be comparing unlike things.
So we define each metric below in terms you can compute yourself with a spreadsheet. The wider measurement approach is in our overview of how to measure GEO performance. Here we focus on the arithmetic.
The raw material: a fixed prompt set and saved answers
Every metric starts from the same table. Each row is one answer: a prompt, an engine, a run number, a date, the full answer text, the brands named in it and the URLs cited in it. Without this table, a percentage cannot be checked by anyone.
The prompt set is a frozen list of buyer questions. Build it using the method in our guide to researching buyer prompts, and do not change it between reporting periods. Add a second set if you need new questions, and report the two sets separately.
Define the words you will use before you count:
- Valid answer. An answer that was returned and saved in full. Exclude errors, refusals and blank runs, and report how many you excluded.
- Mention. The brand name, or an agreed set of spellings and product names, appears in the answer text.
- Citation. The answer shows a link, footnote or source card that points to a URL on a given domain.
- Tracked brand set. Your brand plus a fixed list of competitors you decided on in advance.
The core metrics, defined precisely
Here are five metrics, each with a formula. Let N be the number of valid answers in the slice you are reporting, for example one engine across all prompts and runs.
| Metric | Formula | What it tells you | Common trap |
|---|---|---|---|
| Mention rate | Valid answers that name your brand divided by N | How often the engine brings you up at all | Counting a name that appears only in a list of sources, not in the answer text |
| Citation rate | Valid answers with at least one link to your domain divided by N | How often your pages are used as a source | Counting links to other domains you own without saying so |
| Citation share | Citations to your domain divided by all citations in the same answers | Your slice of the sources the engine relies on | Letting one answer with 15 links outweigh ten answers with two links each |
| Share of voice | Your brand mentions divided by mentions of all brands in the tracked set | Your slice of the brand names the engine produces | Changing the competitor list between periods |
| Answer accuracy | Mentions with correct facts divided by all mentions | Whether what is said about you is right | Judging accuracy from memory instead of a written checklist |
Notice that mention rate and share of voice have different denominators. Mention rate uses answers. Share of voice uses brand mentions. If an engine names six brands in every answer and you are named in half of them, your mention rate is 50 percent but your share of voice is low, because six brands share the space. Both facts are useful, and neither replaces the other.
Citation share has two versions
You can compute citation share by counting every link, or by counting each domain once per answer. The second version is usually fairer, because it stops a single answer that lists many links from dominating. Pick one, write it down and keep it. We prefer once per domain per answer, and we state that in every report.
Prominence is a separate measure
Being first in a list differs from being fourth. If you want to track this, add a simple prominence field: named first, named in a list, mentioned in passing, or only cited as a source. Report counts for each. Do not collapse it into a single weighted score unless you can explain every weight, since invented weights make a number look more scientific than it is.
Mention rate
- Counts answers
- Divide by valid answers
- How often the engine brings you up
Share of voice
- Counts brand mentions
- Divide by all tracked mentions
- Your slice against named competitors
A worked example with illustrative numbers
The numbers below are made up to show the arithmetic. They are not data from any client or engine.
Suppose your prompt set has 20 prompts, you run each three times on one engine, and all 60 answers are valid. Your brand is named in 15 answers. Your domain is cited in 9 answers. Across all 60 answers the engine shows 180 links in total, and 12 of them point to your domain. Three competitors you chose in advance are named 20, 14 and 10 times, and you are named 15 times.
- Mention rate: 15 divided by 60 is 25 percent.
- Citation rate: 9 divided by 60 is 15 percent.
- Citation share by raw links: 12 divided by 180 is about 6.7 percent.
- Share of voice: 15 divided by 59 total tracked mentions (15 + 20 + 14 + 10) is about 25.4 percent.
Those four figures describe four different things. A reader who only sees 25 percent might assume you appear in a quarter of all brand mentions and a quarter of all answers. In this example both happen to be close to 25 percent, but they would drift apart as soon as the number of competitors named per answer changed.
Four metrics from one made up set
Non determinism: why the same prompt gives different answers
AI answers are generated, not looked up. A language model chooses words with some randomness, and many engines also fetch live web results, which change. Two runs of one prompt a minute apart can name different brands. Settings, the user's location, the conversation history, the engine version and the time of day can all shift the output.
This has three practical consequences. First, one run per prompt is a snapshot of one random draw, so repeat each prompt. Second, report the number of runs next to every figure. Third, expect your own numbers to wobble from month to month even if nothing real has changed, and set your threshold for calling something a change accordingly.
Run each prompt at least three times per engine, in the same way, and keep the raw answers. A percentage with no saved answers behind it cannot be audited.
Sample size and how much to trust a percentage
A rate computed from a small number of answers carries a wide margin of error. As a rough guide using the standard formula for a proportion, a rate of 25 percent from 60 independent answers has a 95 percent margin of about 11 points either side. From 300 independent answers it is about 5 points. Those are ranges you should expect from sampling alone.
Real data is worse than that formula suggests, because the formula assumes each answer is independent. Three runs of the same prompt are not independent: if a prompt rarely names you, all three runs will usually agree. So the effective sample is closer to the number of distinct prompts than to the number of runs. With 20 prompts, a change of one prompt moving in or out is 5 points of movement.
| Situation | What you can reasonably say |
|---|---|
| One engine, one run per prompt, 20 prompts | A rough snapshot only. Do not report percentage point changes. |
| One engine, three runs per prompt, 20 prompts | A usable baseline. Report ranges and name which prompts changed. |
| Five engines, three runs per prompt, 20 prompts | A stronger view per engine, but still a sample of 20 questions. |
| Any set, change repeats for two or three periods | A more credible trend, though still not proof of cause. |
Sample design and confidence
Practical rules for clean metrics
- Write the formula for each metric in the report itself, in one line.
- Freeze the prompt set, the tracked brand set and the spelling variants before the first run.
- Report engines separately first, then overall. Never show only a blended figure.
- Show N, the number of valid answers, beside every percentage.
- Keep all raw answers in a dated folder or sheet.
- Restate past periods if you ever change a definition.
- Separate branded prompts, which name you, from unbranded prompts, which do not. Branded prompts will always show a high mention rate and should not be mixed in with discovery prompts.
If you use a tracking product instead of doing this by hand, ask how it computes each of these before you trust its dashboard. Our guide to types of AI visibility tools lists the questions worth asking. And once visibility is measured, connect it to visits with AI search traffic in Google Analytics.
Rules for numbers you can audit
- ✓Write the formula in the report
- ✓Freeze the prompt set and brand set
- ✓Report engines separately first
- ✓Show N beside every percentage
- ✓Keep all raw answers in a dated folder
- Rely on one AI visibility score out of 100
- Average position across engines
Beyond share: accuracy, sentiment and source quality
Share metrics say how often you appear. They do not say whether the answer is right or whether it helps you. Add three quality measures, and keep them as counts you can audit.
| Measure | How to record it | What to do with it |
|---|---|---|
| Accuracy rate | For each answer that names you, mark the facts as correct, partly wrong or wrong, using your approved fact sheet | Wrong facts go to the correction list. See how to fix wrong information |
| Framing | Mark whether you are described as a leader, an option, a budget choice, a niche player, or with a warning | Look for patterns by prompt type. Framing often traces back to third party pages |
| Source type | For each cited URL, tag it: your site, competitor, directory, review site, news, community, reference | Shows where the engine looks. If communities dominate, read our community note |
| Intent fit | Tag each prompt as informational, comparison or ready to buy, and report metrics by tag | Being named only for informational prompts and never for buying prompts is a gap worth fixing |
Avoid turning these into a single sentiment score from a tool that you cannot inspect. A person reading 50 answers and tagging them reaches a more reliable view than an opaque number, and the tags tell you what to do next.
How to choose the competitor set for share of voice
Share of voice depends heavily on who you include. Add ten competitors and your share shrinks without anything real changing. So choose the set once, with a reason, and keep it fixed.
- List the companies buyers actually compare you with, based on sales calls and on which names appear in your test answers.
- Include three to six names. Fewer than three gives a flattering picture. More than eight dilutes every figure.
- Mix sizes. If you are a small firm, include a larger rival and a similar sized one.
- Write each name with its common spellings, product names and abbreviations.
- Review the set once a year, or when a new rival begins to appear often. When you change it, restate earlier periods with the new set or report a break in the series.
Building these metrics in a spreadsheet, step by step
- Make a table called Answers with one row per saved answer. Columns: ID, date, engine, prompt ID, run, valid (1 or 0), brand named (1 or 0), domain cited (1 or 0), and one column per competitor with 1 or 0 for named.
- Add a Links table with one row per cited URL: answer ID, domain, URL, source type.
- Mention rate: count rows where valid is 1 and brand named is 1, divided by rows where valid is 1. A pivot table by engine does this in seconds.
- Citation rate: the same, using domain cited.
- Share of voice: your brand named count divided by the sum of named counts across your brand and the tracked competitors.
- Citation share: count answers where each domain appears at least once, then divide yours by the total across all domains. This is the once per domain per answer version.
- Add the sample size beside every figure, and a margin of error where you can.
This takes less than a day to set up and has no tool cost. If you later buy a tracking product, you can check that its numbers agree with yours on a few prompts.
Metrics that sound useful but mislead
- A single AI visibility score out of 100. It hides the formula and the weights, and you cannot act on it.
- Average position across engines. Positions mean different things in different interfaces.
- Total mentions with no denominator. More prompts produce more mentions, so growth can come from testing more.
- Sentiment percentages with no sample shown. Small samples swing wildly.
- Share of voice against hundreds of brands. The long tail makes every figure tiny and unstable.
- Comparison against a competitor's tool dashboard that you cannot verify.
Turning metrics into actions
| If you see | Likely cause to check | Action |
|---|---|---|
| Low mention rate, low citation rate on all engines | Crawlers blocked, pages thin, weak entity signals | Run the AI Crawler Access Checker, then audit key pages |
| High mention rate, low citation rate | Engine knows you from elsewhere; your pages are not quotable | Rewrite priority pages with direct answers and clear facts |
| High citation rate, low mention rate | Brand name absent from the cited facts | Put the brand next to the claims in headings and opening lines |
| Good on one engine, poor on another | Engine specific retrieval differences | Read the platform page, such as Copilot and Bing, and test its specific sources |
| Share of voice falling while mention rate is flat | Competitors named more often, or your set changed | Check the competitor set, then check what competitors gained |
What to do next
Pick one engine, 10 prompts and three runs each, and compute mention rate and citation rate by hand this week. It takes an afternoon and teaches you more about your visibility than any single dashboard number. The free AI Visibility Checker can give you a first view if you want a starting point.
On the GEO Pilot, a fixed $999 project delivered inside 30 days, we run this exact method in week 1 and again in week 4 on the same prompts, so the before and after comparison uses identical definitions. We report the sample size, we do not promise a particular result, and we say plainly where the data is thin.
