How to Measure AI Visibility: 4 Core Metrics
To measure AI visibility, run a fixed set of buyer-style prompts across AI engines on a schedule and track four metrics: mention rate, share of voice, citations, and sentiment. Because AI answers vary from run to run, repeat each prompt several times and report percentages, not single results. Track the same prompts weekly so you can see trends instead of noise.
Why AI visibility needs its own metrics
Traditional SEO has rankings: you are position 3 or you are not. AI answers don't work that way. An engine might name your brand in one response, skip it in the next, and cite a competitor's page as the source for both.
Most published guides on this topic converge on a similar set of measures: how often you're mentioned, how you compare with competitors, which sources get cited, and how you're described [1][2][5]. The academic work that introduced GEO also treats visibility as something measured inside generated responses, not as a position on a results page [14].
That's the core idea: you're measuring presence, comparison, sourcing, and framing within answers.
Step 1: Build a prompt set before you measure anything
Metrics are only as good as the prompts behind them. Build a list of questions your buyers might ask an AI assistant, and keep it stable.
Group prompts by intent:
- Category discovery: "What are the best [category] tools for [use case]?"
- Comparison: "[Your brand] vs [competitor]: which is better for [audience]?"
- Problem-led: "How do I [solve problem] without [constraint]?"
- Brand-specific: "Is [your brand] good for [use case]?"
Keep branded and non-branded prompts separate in reporting. Branded prompts will almost always show you; non-branded prompts show whether you're earning visibility among people who don't know you yet.
Once the set is locked, change it deliberately (and note the date) rather than editing it casually. Otherwise week-over-week changes may reflect your prompt edits, not real movement.
Step 2: Define the four core metrics
The formulas below are one common way to define these metrics. Tools vary in the details [3][4][6], so write your definitions down and stay consistent.
Mention rate
How often your brand appears in AI answers.
Mention rate = (answers that mention your brand ÷ total answers collected) × 100
This is your baseline metric. Calculate it overall, per engine, and per prompt group. Decide up front what counts as a mention (brand name, product name, common misspellings) and apply the rule the same way every time.
Share of voice
How much of the conversation you own compared with competitors.
Share of voice = (your brand mentions ÷ total mentions of all tracked brands) × 100
Mention rate tells you how visible you are. Share of voice tells you how visible you are relative to the set you chose to compare against. That last part matters: add or remove a competitor and your share changes even if nothing else did. Fix the competitor list and note any changes.
Citations
How often AI engines link to or reference your content as a source.
Two versions are worth tracking:
Citation rate = (answers that cite your domain ÷ total answers collected) × 100
Citation share = (citations to your domain ÷ total citations across all domains) × 100
A mention means the model talked about you. A citation means it pointed to your page as evidence. They don't always move together: you can be mentioned without being cited (the model "knows" you from elsewhere), or cited without being named in the answer text. Also log which URLs get cited, because that tells you which pages are doing the work, and which third-party sources (review sites, forums, publications) are shaping answers about you.
Not every engine shows sources the same way, so compare citation metrics within an engine before comparing across engines.
Sentiment
How you're described when you do appear.
A simple approach is to label each mention as positive, neutral, or negative, then calculate:
Net sentiment = ((positive mentions − negative mentions) ÷ total mentions) × 100
Labeling can be automated or done by hand on a sample, but spot-check it either way. Also capture qualitative notes: Is the model recommending you first? Does it attach a caveat like "better for small teams only"? Those recurring phrases are often more actionable than the score.
Worked example (hypothetical numbers)
Example only. These numbers are invented to show the math. They are not benchmarks.
Say you track 20 prompts on one engine, running each prompt 5 times. That's 100 answers.
| Brand | Answers mentioning brand | Share of total mentions |
|---|---|---|
| Your brand | 34 | 22.7% (34 ÷ 150) |
| Competitor A | 52 | 34.7% |
| Competitor B | 41 | 27.3% |
| Competitor C | 23 | 15.3% |
| Total mentions | 150 | 100% |
From this:
- Mention rate: 34 ÷ 100 = 34%
- Share of voice: 34 ÷ 150 = 22.7%
- Citation rate: your domain appears as a source in 18 of the 100 answers, so 18%
- Sentiment: of your 34 mentions, 20 are positive, 11 neutral, 3 negative. Net sentiment = (20 − 3) ÷ 34 = 50%
What would you do with this? You're mentioned in about a third of answers but cited in under a fifth. That gap suggests models know your brand but don't see your own pages as the best evidence. A reasonable next step is to look at which prompts produce mentions without citations and check whether you have a clear, well-structured page answering those questions.
Step 3: Sample properly by running prompts repeatedly
AI answers are not deterministic. The same prompt can produce different brand lists, different sources, and different wording from one run to the next. A single check is an anecdote, not a measurement.
Practical sampling rules:
- Run each prompt multiple times per cycle. In the example above, 5 runs per prompt turned 20 prompts into 100 data points. More runs reduce the noise in each prompt's result.
- Track across several engines. Different models draw on different data and behave differently, so report each engine separately before blending into a total.
- Keep conditions consistent. Use the same prompt wording, the same settings, and the same cadence each week.
- Report rates, not yes/no. "Mentioned in 2 of 5 runs" (40%) is more honest than "we showed up."
- Look at trends, not single weeks. A one-week dip of a few points can be noise. A sustained multi-week change is a signal.
- Annotate changes. Log content launches, PR hits, site changes, and prompt-set edits so you can connect movement to causes.
Larger prompt sets and more runs give steadier numbers, but they cost more to collect. Dedicated trackers exist largely to automate this repetition [7][9][11].
Step 4: A simple weekly reporting template
Keep the weekly report short enough that people read it. Here's a template you can copy.
| Section | What to include |
|---|---|
| Scope | Number of prompts, runs per prompt, engines tracked, total answers collected |
| Headline metrics | Mention rate, share of voice, citation rate, net sentiment, each with change vs. last week and vs. 4 weeks ago |
| By engine | The same four metrics split per engine |
| By prompt group | Discovery, comparison, problem-led, branded |
| Competitor view | Share of voice for each tracked competitor |
| Top cited sources | Your most-cited URLs and the top third-party domains cited in your category |
| Notable answers | 2–3 quotes or screenshots, especially negative or inaccurate ones |
| Changes shipped | Content published, pages updated, outreach done |
| Next week's actions | 2–3 specific tasks tied to a gap in the data |
A quick checklist before you send it:
- [ ] Same prompt set as last week (or changes noted)
- [ ] Same competitor list
- [ ] Rates shown with sample size
- [ ] Movement compared against a multi-week trend
- [ ] At least one action tied to each major gap
Common mistakes to avoid
- Checking once and calling it a result. Variation between runs is normal.
- Only tracking branded prompts. You'll look great and learn little.
- Changing the competitor set mid-stream. It silently rewrites share of voice.
- Treating mentions and citations as the same thing. They point to different fixes.
- Ignoring inaccuracies. If a model states something wrong about your product, that's a content and sourcing problem worth fixing, even if sentiment looks fine.
Putting it into practice
You can run this process in a spreadsheet to start: a prompt list, a run log, and a few formulas. It gets tedious fast once you add engines and repeat runs.
If you'd rather automate it, growskey.ai tracks ChatGPT, Claude, Gemini, DeepSeek, and a web-search engine. Plans start at $49/month for 50 prompts weekly, and scale up to 500 prompts weekly on the Scale plan. Team members are unlimited, so content, SEO, and PR can all see the same numbers. When the data points to a gap, AI Create can draft GEO articles, comparison pages, and FAQ pages to address it, and post performance tracking helps you see what happened after publishing.
Whatever tool you use, the discipline matters more than the dashboard: stable prompts, repeated runs, consistent definitions, and a weekly habit of turning numbers into actions.
FAQ
What is the most important AI visibility metric?
There isn't one. Mention rate is the best starting point because it's simple and shows whether you appear at all. Pair it with share of voice for competitive context and citations to see whether your content is being used as evidence.
How many times should I run each prompt?
Enough to smooth out variation, and more is better. Many teams run each prompt several times per cycle (the example above uses five). Pick a number you can sustain, keep it constant, and report results as percentages with the sample size.
Is a mention the same as a citation?
No. A mention means the AI named your brand in its answer. A citation means it referenced your content or domain as a source. You can have one without the other, so track them separately.
How often should I report on AI visibility?
Weekly collection is a good cadence for spotting trends, and a monthly summary works well for leadership. Avoid reacting to a single week's change; look for movement that holds across several weeks.
Can I compare results across different AI engines?
Compare trends within each engine first. Engines differ in how they retrieve information and display sources, so absolute numbers aren't always comparable. A blended total is useful as a summary, but keep the per-engine view in your report.
Sources
- AI Visibility Metrics
- AI Visibility Metrics and KPIs: What to Track and Report
- How to Measure AI Visibility | Yext
- AI Visibility Metrics: What to Measure (and How) · Peekl
- GEO KPIs: How To Measure The Right GEO Metrics | Similarweb
- AI Visibility Metrics: The Complete Guide to Measuring Your Brand in AI Engines - CitedSpy
- Which are the Best LLM Visibility Trackers? | Connective3
- Ahrefs Brand Radar for AI Visibility | Capconvert
- AI Search Visibility Tracking and Monitoring | AirOps
- GEO: Generative Engine Optimization
See where your brand stands today
Run a free visibility check: live answers from ChatGPT, Claude and Gemini to the questions your buyers ask.
Check visibility