AI visibility is how often AI systems such as ChatGPT and Google's AI Overviews name your brand or cite your pages when people ask the questions your buyers ask. You measure it by putting a fixed set of those questions to each system, several times over, and counting what comes back. A single check gives you a single number, and that number can mislead, because the same question gets a different answer each time.
Suppose ChatGPT names you in 30% of its answers to a question. Check once, and 7 times out of 10 you'll be told you're invisible. Check again next week with nothing changed, and the two results will disagree 42% of the time. A dashboard built on single checks can show you movement that isn't there.
I checked five explainer guides from Google's first two pages of results for "ai visibility" on 18 September 2026 (US, desktop). Three recommend a single visibility score. None puts a range on a number or says when a change counts as real. This guide covers those two things.
Find out what the answers say about you.
One measurement, your own crawler data, and the three access findings that matter most. Free, no card, about two minutes.
Start freeWhat is AI visibility?
AI visibility is the share of relevant AI answers in which your brand appears, and how it appears. There's no position to hold, because an answer is written fresh each time. Four counts describe it:
| What you count | The question it answers |
|---|---|
| Mention rate | How often are you named at all? |
| Share of voice | Of all the brand names in those answers, what share is yours? |
| Prominence | When you're named, how early in the answer? |
| Citation share | How often is your site one of the linked sources? |
The four can move separately. An answer can recommend you without linking to you, and it can cite your page while recommending a competitor. Share of voice also rises when a competitor falls, so read it next to mention rate. Shruwd, the tool we are building, documents the exact definition of each.
Why one number lies
The same question gets different answers
SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times in late 2025. One of their findings: "there's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses." The volunteers used their own accounts and locations, so the figure mixes randomness with personal differences.
An April 2026 preprint titled "Don't Measure Once" puts it in one line: "Answers can vary across runs, prompts, and time, making one-off observations unreliable."
Semrush's head of AI visibility advises reading every number as a range:
a share of voice that swings between 20% and 40% over a day is normal, so "30% ± 10%" is the honest way to report it.
Small samples give wide ranges
Say you're named in a third of the answers you collect. How much that tells you depends on how many answers there were:
| Answers collected | Named in | Rate | 95% range |
|---|---|---|---|
| 3 | 1 | 33% | 6% to 79% |
| 9 | 3 | 33% | 12% to 65% |
| 30 | 10 | 33% | 19% to 51% |
| 90 | 30 | 33% | 24% to 44% |
| 300 | 100 | 33% | 28% to 39% |
The rate is 33% in every row. The first two rows tell you almost nothing. The range is where the real rate probably sits, given the answers collected. Statisticians call it a 95% confidence interval, and Shruwd's docs explain why every number carries one.
The score depends on the questions you picked
An AI share of voice of 40% means 40% across the questions someone chose. Change the list and the number changes. Dan Taylor made the point in Search Engine Land: "these metrics rely on a hidden denominator." He contrasts it with traditional search, "where visibility could be measured against a known keyword set", while "the universe of possible AI prompts is effectively infinite."
So you can't compare a percentage with another company's number, or with your own from a different list.
Answers change even when you don't
Profound compared the sources AI answers cited in June and July 2025, across about 80,000 prompts per platform. Of the domains cited in July, 59.3% hadn't been cited for the same prompt in June on Google AI Overviews. The figure was 54.1% on ChatGPT and 40.5% on Perplexity. There was no same-day baseline, so part of that drift is the ordinary run-to-run variation described above.
The companies behind the engines also update their models. A drop in the week a model changes may have nothing to do with your site.
How to measure AI visibility
1. Fix a list of questions
Write 20 to 50 questions a buyer would ask, in their words, weighted toward buying and comparison questions. Then leave the list alone. Repeating the same questions is what makes this month comparable with last month. Shruwd's docs have more on choosing prompts.
2. Ask each question several times, on each engine
After their study, the SparkToro authors concluded that "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric."
What counts is the total number of answers behind each number. With 25 questions asked three times a week, you collect 75 answers per engine per week. If a third of them name you, the range is 24% to 45%. After four weeks and 300 answers it's 28% to 39%. A single question asked three times a week needs four weeks to reach a dozen answers, so judge the list as a whole first.
3. Count the four things separately
Some tools blend several measures into one AI visibility score. When a blend drops, you can't tell which part moved. Keep mention rate, share of voice, prominence and citation share apart.
4. Put a range on every number, and hold back the thin ones
To get a range yourself, put two numbers into any Wilson score interval calculator: how many answers you collected, and how many named you. A rate from six answers carries almost no information. Shruwd withholds any number with fewer than ten responses behind it and shows "insufficient data", because a printed number invites a decision.
5. Decide what counts as a change before you look
Two rules, both required. The ranges before and after must not overlap, and the change must be at least five percentage points. With 75 answers on each side, a mention rate of 33% has to reach about 56% before the ranges separate. With 300 on each side, about 45% is enough.
Watching many questions creates its own trap. Test 50 questions, each with a 5% chance of a false alarm, and you should expect 2.5 false alarms every round. A correction for testing many questions at once keeps those in check. Shruwd's docs set out the full rule, including how it marks a comparison where the model changed in between.
What is a good AI visibility score?
There isn't a universal one. A score depends on the questions chosen, the engines, the competitors tracked and how the tool blends its parts, so 62 in one tool and 62 in another describe different things.
Two comparisons are worth making. One is your mention rate against your competitors' on the same questions. The other is your own rate against last month's. Both need ranges.
What the answers can't show you: traffic
Counting answers tells you whether you're recommended. It doesn't tell you whether anyone arrived.
Expect small numbers. Conductor measured AI referrals at 1.08% of website traffic across 13,770 domains in ten industries, in the US between May and September 2025.
Google Analytics has reported that traffic natively since 13 May 2026, when it added an AI Assistant channel to the default channel group. It sorts visits by the site they came from, and covers ChatGPT, Gemini, Copilot and others. It "excludes Google's AI Overviews and AI Mode", so for those use the Generative AI performance report in Search Console, which Google's guide points to.
Both are likely to undercount. A buyer who sees your name in ChatGPT and searches for you the next day arrives as branded search.
Questions to ask any AI visibility tracker
- How many times is each question asked per period?
- Does every number come with a range, and what is the smallest sample it will report?
- Is there a blended score, and can you see its parts?
- Do the answers come from an API or from the app a person uses, and for which location?
- What rule decides that a number has moved?
- Can it tell you why you're missing, or only how often?
Google's guide has a line for every vendor in this market, me included: "No third-party tool has access to our internal ranking or AI systems."
Here are Shruwd's answers. On the paid plans the questions run weekly, and each one is asked three times per run on each engine. Every rate carries a 95% range, and no number is shown with fewer than ten responses behind it. There's no blended score. Movement follows the two rules above, with the correction for many questions. The free plan takes one snapshot of ten questions asked once each, which is the smallest sample Shruwd will put a number on. The plan table says so.
The answers come from two places. Google AI Overviews are read from Google's results pages, fetched for one set location. Your buyers search from many places, so that's a stand-in too. ChatGPT answers come from OpenAI's model through an API with web search switched on. That isn't the ChatGPT app. There's no memory, chat history or personalisation.
Surfer compared API answers with app answers over 1,000 runs each and found only 24% of brands in common. Its test had no baseline for how much two app runs differ from each other, so some of that gap is ordinary variation. Some of it is probably real. Read Shruwd's ChatGPT numbers as a stand-in for the app, collected the same way each week. That makes them more useful for tracking change over time than as a picture of what one signed-in person sees.
On the last question, Shruwd reports likely reasons as findings, each with one of three confidence labels. The findings reference lists them.
Frequently asked questions
How can I check my AI visibility?
Write down ten questions a buyer would ask. Ask each one in ChatGPT and in Google five times, in fresh sessions, and note every brand named and every site linked. That's up to fifty answers per engine, fewer where Google shows no AI answer. Fifty answers gives a range of roughly 19% to 44% around a rate of 30%. That's enough to tell whether you're named often, sometimes or almost never. It won't detect a small change.
How do you increase AI visibility?
Check the mechanical things first: that AI crawlers can reach your pages, and that your content is in the raw HTML. Then work on being mentioned on the pages that answers cite. LLM SEO: five levers with evidence grades each tactic by the evidence behind it, and GEO vs SEO covers what changed and what didn't.
Does Google Analytics show AI traffic?
Yes, since 13 May 2026. The default channel group has an AI Assistant channel for visits referred by ChatGPT, Gemini, Copilot and similar tools. It leaves out Google's AI Overviews and AI Mode. ChatGPT also adds utm_source=chatgpt.com to the links it sends, according to OpenAI's publisher FAQ.
