How to measure AI visibility: mention rate, citation share and share of voice
Measure AI visibility as a rate across many runs of a fixed prompt set, not as a ranking position. Track how often you are named, how often your pages are cited and how you compare with competitors, then add the free data from Search Console, Bing and GA4.
In short
- Mention rate is the share of answers to a fixed prompt set that name your brand, and it is the one metric that SparkToro's January 2026 research found stable across repeated runs.
- In that research, 2,961 runs of 12 prompts returned the same list of brands less than 1% of the time, so a single "AI ranking position" has no meaning.
- Mentions and citations must be counted separately, because Semrush's 2026 AI Visibility Index found their overlap on Gemini was as low as 30%.
- Bing Webmaster Tools has shown Copilot citations and the grounding queries behind them since February 2026, at no cost, with the query data given as a sample.
- A working guide from practitioners is 50 to 300 prompts per market, each run many times, reported monthly with ranges.
You track AI visibility by asking a fixed set of questions many times and counting how often your brand appears. The result is a rate, such as “named in 34% of answers”, and never a position. Alongside that, you read the first-party reports that Google and Microsoft now provide, which show a different part of the picture. The rest of this lesson defines the metrics, shows how to build the prompt set and explains what each free data source can and cannot tell you, because the choices you make at the start decide whether the numbers mean anything six months later. The tools that automate the work are covered in AI visibility tracking tools.
Why a position does not work
A search result page is fairly stable, which is why rank tracking has worked for so long: you look once, write down the position, and the number is still broadly true the next day. An AI answer is written anew each time, so the same habit breaks. SparkToro and Gumshoe published a test on 30 January 2026 in which 600 volunteers ran 12 prompts 2,961 times across ChatGPT, Claude and Google’s AI. The same list of brands came back less than 1% of the time, and the same list in the same order less than 0.1% of the time. If the order repeats that rarely, a report that says “we are second in ChatGPT for this prompt” describes one answer that happened once, and the next person to ask will see something else.
What did hold in that research was frequency. Leading headphone brands appeared in 55% to 77% of responses even when the wording varied. That finding is what makes measurement possible at all, because it means a brand has a fairly steady chance of being named even though any single answer is unpredictable. So the unit of measurement is how often you appear across many runs, and every metric below is some version of that count. It also follows that a figure taken from a handful of runs is weak, since a brand that appears in roughly half of all answers can easily show up in two runs out of five or four out of five by chance alone.
The core metrics
Seven measures cover most of what a team needs to report. The first four come from prompt testing, the fifth needs a person to read the answers, and the last two come from your own analytics and search data.
| Metric | What it counts | What it is for |
|---|---|---|
| Mention rate | Share of answers that name your brand | The headline number |
| AI share of voice | Your mentions divided by all tracked brands’ mentions | Comparison with a fixed competitor set |
| Citation rate | Share of answers that link to your domain | Whether your own pages are retrieved |
| Third-party source coverage | Of the pages cited for your prompts, the share that mention you | The off-page to-do list |
| Sentiment and accuracy | How the brand is described, and whether it is correct | Needs a person to read answers |
| AI referral sessions | Visits arriving from assistant domains | Undercounts influence |
| Branded search trend | Searches for your name over time | A proxy for influence without clicks |
The table needs three notes before you rely on it. The first is that mentions and citations have to be counted separately. A mention is your name in the answer text, while a citation is your URL as a source, and one can appear without the other. Semrush’s 2026 AI Visibility Index found the overlap between mentioned brands and cited domains on Gemini was as low as 30%. Semrush sells a tracking product, so treat that as a vendor figure, but the direction of it matters for how you report: a single blended “visibility” number would hide the case where an assistant recommends you without linking to you, and the opposite case where it uses your page as a source and recommends someone else.
The second note is that share of voice depends on your list. Add a large competitor and your share falls without anything changing in the answers, simply because the total you are divided by has grown. That is why the competitor set should be fixed at the start and any change to it written down with a date, so that a later reader does not mistake an edit to the list for a loss of ground. The third note is that third-party source coverage is the metric you can act on. If 40 pages are cited repeatedly for your prompts and 6 of them mention you, the other 34 are a target list, and each one is a page where an editor, a reviewer or a community could be given a reason to include you. That method is set out in link building for AI visibility.
A worked example makes the arithmetic plain, and the figures in it are invented for that purpose only. Say you track 100 prompts, run each 10 times on one assistant, and so collect 1,000 answers. Your brand is named in 280 of those answers, so the mention rate is 28%. Across the same answers, all tracked brands are named 2,000 times in total and yours accounts for 280 of those, so share of voice is 14%. Your domain is linked in 90 answers, which gives a citation rate of 9%. Notice that the three numbers come from the same 1,000 answers and still differ widely, which is the reason for keeping them apart. Report each figure per assistant as well. Combining assistants into one score hides the fact that they retrieve from different indexes, so a strong result on one can mask a blank on another.
Building the prompt set
No one outside the platforms knows how often real people ask a given prompt. The “AI search volume” figures in tools are modelled. So you choose prompts by judgement, and the choice shapes every number that follows: a set full of questions where you are already well known will report a comfortable rate, and a set built from the questions buyers ask before they have heard of you will report a much less comfortable one. Neither is wrong, but only the second tells you where the work is. A working method has five steps, and the order matters because the later steps depend on the earlier ones being settled.
- Start from buyers’ questions, drawn from sales calls, support tickets, site search and “people also ask” boxes.
- Cover the stages: category questions (“best payroll software for small firms”), comparison questions (“A or B”) and brand questions (“is A reliable”).
- Keep brand prompts apart. A prompt that contains your name will nearly always mention you, so report those separately or they flatter the rate.
- Fix the set, and add prompts only in dated batches. A rate from this month’s set cannot be compared with a rate from a different set.
- Run each prompt many times. A working guide is 50 to 300 prompts per market.
That range and the practice of repeated runs are rules of thumb from practitioners, not tested thresholds, so treat them as a sensible starting point and not as a standard anyone has validated. The reasoning behind them is sound all the same. More prompts spread the measurement across more of the questions buyers ask, and more runs per prompt reduce the part of each figure that is down to chance. If the budget forces a choice, it is better to say plainly how many runs sit behind a number than to stretch a thin sample across a long prompt list and present the result as precise.
Free first-party data
Four sources cost nothing and each sees a different part of the picture. In Google Search Console, AI Overviews and AI Mode have counted in the Web search totals since 17 June 2025, which means impressions and clicks from those surfaces are already inside the figures you look at. Google announced dedicated generative AI performance reports on 3 June 2026 and rolled them out in stages. We have not confirmed in a live account which breakdowns they include, so check your own property for whether clicks and queries are shown. Bing Webmaster Tools goes further for Microsoft’s surfaces. Microsoft’s blog post of 10 February 2026 introduced an AI Performance report that shows total citations, average cited pages per day, citations by page and grounding queries, which Microsoft defines as “the key phrases the AI used when retrieving content”. It covers Copilot, AI summaries in Bing and some partner products. Microsoft describes the grounding query data as a sample, and the report has no click data, so it tells you which pages are being used as sources without telling you whether anyone followed the link.
The other two sources sit on your own side. Google’s Analytics help page lists an “AI Assistant” default channel in GA4 for visits from sources such as ChatGPT, Gemini, Copilot and Grok, while clicks from AI Overviews and AI Mode stay in Organic Search; how much traffic AI search sends explains the set-up. Server logs add a view that analytics cannot give. OpenAI’s crawler documentation describes a ChatGPT-User agent that fetches pages when a user’s request needs them, and similar agents exist for Perplexity and Claude, so their hits show which of your pages assistants load. None of these four sources tells you whether ChatGPT names your brand in its answer. That gap is why prompt testing exists, and it is why the free data works best as a companion to a prompt set and not as a replacement for one.
Limits and a reporting routine
Some limits belong in every report, because a reader who does not know them will read more certainty into the figures than they hold. Answers vary with location, memory, model version and mode, so your tests see one slice of what real users see. A tool that queries a model through its API may not get the answer a person sees in the app, which makes it worth asking any vendor which of the two it samples. Assistants also rewrite prompts into hidden searches, a behaviour explained in query fan-out, so the prompt you track is only the visible start of the retrieval. And the studies that link any action to higher visibility are correlational and mostly published by vendors. A rising mention rate after a campaign is an observation, not proof of cause, and it should be written up that way.
With those limits stated, the routine itself is short. Each month, report mention rate, share of voice and citation rate per assistant, each with the number of runs behind it and the range across runs, so that a reader can tell a real change from ordinary variation. In the same monthly report, list the most-cited third-party pages for your prompts and note which of them now mention you, since that list is where the off-page work shows up first. Each quarter, add a manual read of 20 to 30 answers for accuracy. Wrong facts often trace back to an outdated profile, which is entity work and has to be fixed at the source. Ignore day-to-day movement and any single answer, however striking it looks in a screenshot.
BuzzStream’s State of Digital PR 2026 found that 55.4% of digital PR practitioners track AI citations as a success metric and about 40% have a repeatable method. Set side by side, those two figures point to a group that reports the metric without a consistent way of producing it. A fixed prompt set and a monthly rate is that method, and it needs discipline about keeping the set and the competitor list unchanged far more than it needs expensive software.
Where to go next
The AI visibility tools comparison lists trackers by category, which is the natural next step once the prompt set is written and you know how many prompts and assistants you need to cover. For the campaigns that move these numbers, read digital PR. Metric names used here are defined in the glossary.
Common questions
What is AI share of voice?
It is your brand's mentions divided by all mentions of the brands you track, across the same set of prompts. It needs a fixed competitor list, and it changes when you change that list.
What is the difference between mention rate and citation rate?
Mention rate counts answers that name your brand in the text. Citation rate counts answers that link to your domain as a source. A brand can be named with no link, and a page can be linked without the brand being recommended.
How many prompts should I track?
A working guide is 50 to 300 prompts per market. That is a rule of thumb from practitioners, not a tested threshold. The number of runs per prompt matters as much as the number of prompts.
Can I measure AI visibility for free?
Partly. Search Console counts Google's AI surfaces, Bing Webmaster Tools reports Copilot citations and GA4 has an AI Assistant channel for referral visits. None of them shows whether ChatGPT names your brand, which needs prompt testing.
Why do AI visibility numbers change so much from day to day?
The answers are generated afresh each time and vary with wording, location, model version and memory. Daily movement is mostly noise. Compare monthly rates across many runs.

