AI visibility is measured by asking the same unbranded questions many times and reporting how often answers recommend or cite you, with a Wilson interval. One answer is not a measurement. A rate of 30% from 10 samples, from 100 and from 1,000 means three different things.
- A rate of 30% has a 95% Wilson interval of 10.8% to 60.3% from 10 samples, 21.9% to 39.6% from 100 and 27.2% to 32.9% from 1,000 (Penlid, methodology page, read 30 October 2026).
- Penlid reports two figures as significantly different only when their 95% intervals do not overlap, and it applies no correction for comparing many pairs (Penlid, methodology page, read 30 October 2026).
- A 2026 arXiv preprint on generative search measurement argues that
citation visibility must be reported with uncertainty estimates
(arXiv, Quantifying Uncertainty in AI Visibility, read 30 October 2026). - Penlid's free AI visibility check asks one engine, Claude with live web search, four of the questions your buyers type (Penlid, AI visibility check page, read 30 October 2026).
Why is one AI answer not a measurement?
One AI answer is not a measurement because the same question can return a different answer each time you ask it. Our methodology says Answers from language models are stochastic: ask the same question twice and you can get two different sets of recommended vendors.
OpenAI's text generation guide says the content generated from a model is non-deterministic
, and the arXiv preprint Don't Measure Once puts the consequence this way: Answers can vary across runs, prompts, and time, making one-off observations unreliable.
Measuring visibility therefore means taking many samples and describing how the results spread. Our guide to checking whether ChatGPT recommends your brand applies the idea to one engine.
How do you turn repeated answers into a rate?
You turn repeated answers into a rate by choosing what counts as a hit, asking the same questions many times, and dividing the hits by the number of answers: if k of n answers recommend you, the rate is k ÷ n, and n counts answers, not questions. Count one outcome per rate, because our free check reports two separately: Being listed as a source and being recommended as an option are different outcomes.
Ask the questions a buyer would type without your brand name: We ask the comparison, alternatives and buyer's-guide questions a real customer types.
A question that names your brand all but ensures a mention, so it measures nothing (see why unbranded prompts are the ones worth tracking). Then repeat each question: We sample each tracked prompt repeatedly across each engine on a schedule, and we record the model version returned with every sample.
What does 30% mean at 10, 100 and 1,000 samples?
A rate of 30% means three different things at 10, 100 and 1,000 samples, because the 95% Wilson confidence interval around it narrows from about 50 points wide to about 6. The intervals in the table are from our methodology; the width column is our subtraction of the two printed ends.
| Observed | Samples | 95% Wilson interval | Width |
|---|---|---|---|
| 3 of 10 | 10 | 10.8% to 60.3% | 49.5 points |
| 30 of 100 | 100 | 21.9% to 39.6% | 17.7 points |
| 300 of 1,000 | 1,000 | 27.2% to 32.9% | 5.7 points |
Our methodology says of these three rates, They do not mean the same thing, and we don't display them as though they do.
In plain terms, read each interval as the range of true rates the samples do not rule out at the 95% level.
Ten times the samples narrows the interval about threefold, not tenfold. From 100 to 1,000 samples the width falls from 17.7 to 5.7 points, and 17.7 ÷ 5.7 = 3.1, close to the square root of 10, which is 3.2; the ratio is our arithmetic.
Why use a Wilson interval instead of the textbook formula?
A Wilson interval is used because the textbook normal interval breaks down with few samples and for rates close to 0% or 100%. Our methodology says the normal interval behaves badly at small n and near 0% or 100%, which is exactly where most brands start
. The Wilson interval dates from a 1927 paper (see Sources). The table compares the two at four samples, the size of our free AI visibility check: the Wilson column uses the formula in the example below, and the normal column is p ± 1.959964 × √(p(1 − p) ÷ n).
| Observed | Wilson interval | Normal interval |
|---|---|---|
| 0 of 4 | 0.0% to 49.0% | 0.0% to 0.0% |
| 1 of 4 | 4.6% to 69.9% | −17.4% to 67.4% |
| 3 of 4 | 30.1% to 95.4% | 32.6% to 117.4% |
| 4 of 4 | 51.0% to 100.0% | 100.0% to 100.0% |
For 0 of 4 and 4 of 4 the normal interval has zero width, as if four samples had settled the rate, and for 1 of 4 it is 0.25 ± 0.424, which dips below 0%. The Wilson interval stays inside 0% to 100%, and for 0 of 4 it still reaches 49.0%: four misses cannot rule out a true rate near one in two. The page for our check says its results come with intervals because four samples is four samples and we won't pretend otherwise
.
How many samples do you need?
You need about 80 samples to bring the 95% Wilson interval around a 30% rate to a half-width of 10 points, and about 320 to bring it to a half-width of 5 points. Half-width means half the interval's width, so these intervals are about 20 and 10 points wide in all. Both figures are our calculation with the 95% Wilson formula, not numbers from a source.
| Half-width at most | Samples | Example count | 95% Wilson interval |
|---|---|---|---|
| 10 points | 80 | 24 of 80 | 21.1% to 40.8% |
| 5 points | 320 | 96 of 320 | 25.2% to 35.2% |
For 24 of 80 the interval is 21.1% to 40.8%, which sits 8.9 points below 30% and 10.8 above it; its half-width is (40.8 − 21.1) ÷ 2 = 9.85 points, under the 10-point target. Halving the half-width costs about four times the samples: 80 × 4 = 320. Too few samples can hide a real rate as well: if a brand's true rate were 30%, four samples would show no recommendation at all with probability 0.7 × 0.7 × 0.7 × 0.7 = 0.24, about one time in four.
How do you tell whether two rates really differ?
You can call two rates different when their 95% intervals do not overlap, which is the rule our dashboard applies when it compares one engine against another: Two figures are reported as significantly different only when their intervals do not overlap.
When they overlap, we report no significant change even when the two observed rates differ.
The rule is conservative. Our methodology says two proportions can genuinely differ while their intervals still touch
, so an overlap means the pair is not shown to differ, which is weaker than showing the two are the same.
Comparing many pairs makes a false alarm more likely. Five engines make 5 × 4 ÷ 2 = 10 pairs, and our methodology says testing ten pairs at 95% finds a spurious one more often than one pair does
. The rule applies no correction for this, and when the dashboard reports that two engines differ, it prints how many pairs it compared.
Do not compare across a model change: A rate that spans a model change is two measurements, not one.
An engine can change the model behind its answers without anything on your site changing (see why AI visibility changes when you changed nothing).
Example (illustrative numbers)
These numbers are invented and are not data from any site. A brand asks one engine 7 unbranded questions 10 times each, every week, so a week has n = 7 × 10 = 70 answers. Week 1 has 21 recommendations and week 2 has 35.
The Wilson interval for 21 of 70. For k of n, let p = k ÷ n and z = 1.959964, so z² = 3.8415. The center is (p + z² ÷ (2n)) ÷ (1 + z² ÷ n), and the half-width is z × √(p(1 − p) ÷ n + z² ÷ (4n²)) ÷ (1 + z² ÷ n). Here p = 21 ÷ 70 = 0.30 and z² ÷ n = 3.8415 ÷ 70 = 0.05488, so the center is (0.30 + 0.02744) ÷ 1.05488 = 0.3104. The half-width is 1.959964 × √(0.003 + 0.000196) ÷ 1.05488 = 1.959964 × 0.05653 ÷ 1.05488 = 0.1050. The interval is 0.3104 − 0.1050 = 0.2054 to 0.3104 + 0.1050 = 0.4154, which is 20.5% to 41.5%.
| Answers a week | Week | Recommended | Rate | 95% Wilson interval |
|---|---|---|---|---|
| 70 | 1 | 21 | 30% | 20.5% to 41.5% |
| 70 | 2 | 35 | 50% | 38.6% to 61.4% |
| 100 | 1 | 30 | 30% | 21.9% to 39.6% |
| 100 | 2 | 50 | 50% | 40.4% to 59.6% |
The comparison. At 70 answers a week the intervals overlap, because week 2 starts at 38.6% and week 1 ends at 41.5%, an overlap of 41.5 − 38.6 = 2.9 points. Applied to these two weeks, the rule therefore reports no significant change, although the rate rose from 30% to 50%. A standard two-proportion z test is less strict: the pooled rate is (21 + 35) ÷ (70 + 70) = 0.40, so z = (0.50 − 0.30) ÷ √(0.40 × 0.60 × (1 ÷ 70 + 1 ÷ 70)) = 0.20 ÷ 0.0828 = 2.42, a two-sided p of about 0.016. At 100 answers a week the intervals do not overlap, because 39.6% is below 40.4%, so the rule reports a significant difference. The rates are the same; only the number of samples changed.
What this cannot tell you
An interval measures sampling noise for the questions you asked and nothing else. Our methodology says Prompt selection shapes the result.
and that there is no neutral prompt set, so a different set of questions can give a different rate.
- It cannot tell you why a rate moved.
We observe what an engine answered, not why.
A model update, a change in the engine's retrieval, a competitor's PR win or your own content can each move it (see attributing AI citations to a content change). - It assumes each answer is an independent draw with the same chance of recommending you. That is our reasoning, not a finding from a source: answers to one question on one day may resemble each other more than answers on different days, so treat the sample sizes above as minimums.
- It cannot tell you about visits or sales. That is our reading of the method: it counts what an engine answered and does not see what a reader did next.
Questions people also ask
What is AI visibility?
AI visibility is how often an AI engine's answer recommends your brand or cites your site when someone asks a question your buyers would ask. Because the same question can return a different answer each time, it is a rate over many answers rather than a yes or no. For how it differs from classic search measurement, see our guide to SEO, AEO and GEO.
How do you measure AI search visibility?
Choose questions a buyer would type without your brand name, send each one to an engine many times, and count the answers that recommend you. Divide the hits by the number of answers and attach a 95% interval, so a reader can see how far chance alone could move the rate. Keep the engine's name and the model version beside every rate, and call a gap between two rates significant only when the intervals stay apart, as we do.
What metrics show AI search visibility?
Two rates are worth reporting: how often an answer recommends you and how often it cites your site as a source, each with a 95% Wilson interval, as our free check does. We keep them separate because the two are different outcomes. For Google's own AI-feature reporting, see our guide to the Generative AI performance report.
Is there a tool to check AI visibility?
Penlid's free AI visibility check is one. You enter a domain name, it needs no account, and it asks Claude with live web search four of the questions your buyers type, then shows who gets recommended, with a 95% Wilson interval on your recommended rate and on your cited rate. It samples one engine, and four samples give wide intervals, so treat it as a first look and not as a rate. Whatever tool you try, ask how many times it sampled each question and whether it shows intervals.
Sources
- Methodology: how we measure, and what we can't, Penlid, read 30 October 2026.
Answers from language models are stochastic: ask the same question twice and you can get two different sets of recommended vendors.
We ask the comparison, alternatives and buyer's-guide questions a real customer types.
We sample each tracked prompt repeatedly across each engine on a schedule, and we record the model version returned with every sample.
They do not mean the same thing, and we don't display them as though they do.
the normal interval behaves badly at small n and near 0% or 100%, which is exactly where most brands start
Two figures are reported as significantly different only when their intervals do not overlap.
two proportions can genuinely differ while their intervals still touch
testing ten pairs at 95% finds a spurious one more often than one pair does
A rate that spans a model change is two measurements, not one.
Prompt selection shapes the result.
We observe what an engine answered, not why.
- Free AI Visibility Check: does AI recommend you?, Penlid, read 30 October 2026.
Being listed as a source and being recommended as an option are different outcomes.
because four samples is four samples and we won't pretend otherwise
- Text generation, OpenAI API documentation, OpenAI, read 30 October 2026.
the content generated from a model is non-deterministic
- Don't Measure Once: Measuring Visibility in AI Search (GEO), arXiv, read 30 October 2026.
Answers can vary across runs, prompts, and time, making one-off observations unreliable.
- Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement, arXiv, read 30 October 2026.
citation visibility must be reported with uncertainty estimates
- Probable Inference, the Law of Succession, and Statistical Inference, Journal of the American Statistical Association, vol. 22, no. 158, June 1927, read 30 October 2026.