The central limit theorem says that averages of many independent observations follow an approximately normal distribution, whatever shape the original data had. It is why standard statistical tests work on skewed business metrics without any transformation.
Take any distribution — skewed, bimodal, mostly zeros, it does not matter — draw a sample from it and compute the average. Do that many times and the collection of averages will be approximately normally distributed around the true mean, with a spread that shrinks as the samples get larger. That is the central limit theorem, and it is the reason a body of statistics built on the normal distribution can be applied to data that looks nothing like a bell curve.
It is what licenses the everyday toolkit. A t-test, a z-test, a confidence interval on a conversion rate — all of them assume the *estimate* is approximately normally distributed, and none of them assume the underlying observations are. Revenue per user can be 71% zeros with a long right tail, and the difference in mean revenue between two arms will still be close to normal at experiment-scale sample sizes. Most anxiety about normality in practice is anxiety about the wrong distribution.
The theorem needs three things and it is worth knowing which one usually fails. Independent observations, which is the condition that breaks most often in practice: sessions from one user, weeks from one market, events from one account are all correlated, and correlated observations mean the effective sample is far smaller than the row count. Identical distribution, which can be relaxed considerably. And finite variance, which sounds like a technicality and matters for genuinely heavy-tailed quantities where a single observation can dominate any sample.
How fast the convergence happens depends on the shape of the original data, and the familiar rule of thirty observations is far too optimistic for business metrics. Roughly symmetric data converges quickly, and thirty is genuinely adequate. Heavily right-skewed data with many zeros — which describes most revenue and engagement metrics — may need hundreds or thousands before the mean behaves normally. A metric where the mean is driven by a handful of enormous values may never converge usefully at any sample size you will collect, which is the real reason to winsorise rather than a matter of tidiness.
The theorem also explains the 1/√n that governs the economics of experimentation. The spread of the sampling distribution is the standard error, which falls with the square root of the sample size, so four times the data halves the uncertainty and a hundred times reduces it only tenfold. And it says nothing about statistics that are not averages — a maximum, a ratio of random quantities, a percentile — which is why those need bootstrapping rather than a formula.
The statement, its practical consequence, and the two quantities that follow from it.
( x̄ₙ − μ ) / ( σ/√n ) → N(0, 1) as n → ∞The standardised sample mean converges to the standard normal, whatever distribution the observations came from — provided the variance is finite.
x̄ ~ approximately N( μ, σ²/n )The sampling distribution of the mean is centred on the truth with variance σ²/n. Everything in a results table follows from this line.
SE = σ / √nThe square root is why detecting small effects is expensive: halving the uncertainty costs four times the data.
n_eff = n / ( 1 + (m − 1)·ρ )Independence is the assumption that fails most often. With 5 sessions per user at ρ = 0.35, 200,000 sessions carry about 83,000 observations' worth of information — see the two-sample t-test calculator.
A simulation draws samples from the real distribution of revenue per user — 71% zeros, mean £6.80, standard deviation £91, skewness 14.2 — and records the sample mean each time, at four different sample sizes. Ten thousand repetitions at each size.
At n = 30 the sampling distribution is still visibly skewed and a nominal 95% interval covers only 88% of the time. By n = 30,000 it is effectively normal and the coverage is correct.
The rule of thirty fails badly here, and the coverage column is why it matters: at n = 30 an interval advertised as 95% is really about 88%, so roughly one conclusion in eight is wrong at a rate nobody has budgeted for. The convergence is real but slow — skewness falls roughly as 1/√n, which is why it takes a thousandfold increase in sample to get from 2.6 to 0.08. The practical reading is that experiment-scale samples are comfortably fine for this metric and small pilot analyses on it are not, and that the amount of data needed depends on the skewness rather than on a fixed threshold. Winsorising the top 1% would cut the population skewness by more than half and get the same coverage at a fraction of the sample.
Learn why deviations from normality are often not a significant concern in A/B testing, understand the pros and cons of using the t-test in such scenarios, and explore alternative methods to the traditional t-test.
Many analysts make fundamental mistakes that can misrepresent or undermine the test's findings. This blog highlights common reporting errors and offers practical guidance for presenting tests results effectively.
Applying it to a live measurement problem is the part that goes wrong. If you are designing an experiment, reading a result you do not trust, or trying to work out what your marketing actually caused, that is the work we do.