Tool
Random Sample From a List
Draw a fixed-size sample without replacement — for audits, spot checks, and surveys.
Random Sample From a List
Paste your population above to see the sampling fraction.
Paste a population, say how many you need, and draw exactly that many — no repeats, every item equally likely, and the sampling fraction shown so you know what proportion you are actually looking at.
This is the tool for spot checks and audits rather than for picking a winner. The difference is that a winner is one item and a sample is a claim about the whole population, which means the size of the sample is the part that carries meaning.
It draws without replacement, so nothing can appear twice. That is the right model when you are physically going to inspect each item and inspecting the same invoice twice tells you nothing new.
How the Random Sample From a List works — and why it's fair
The draw is a partial Fisher-Yates pass. Rather than shuffling the whole population and taking the first n, it performs exactly n swaps: pick a random index from the untouched remainder, swap it to the front, move the boundary, repeat. The result is a genuine simple random sample and the work is proportional to n rather than to the population size.
Every subset of size n is equally likely, which is the specific guarantee that makes a sample defensible. That is stronger than every item having an equal chance of being in the sample, though the two go together here — a scheme could give every item equal chance while making certain combinations impossible, and that would fail the test an auditor cares about.
Without replacement is the model, and it matters. Once an item is drawn it is removed from the remaining pool, so the second draw is 1-in-(N−1), the third is 1-in-(N−2), and so on. The odds rise as the sample grows, which is why a 50-of-100 sample is a very different object from fifty independent 1-in-100 draws.
The sampling fraction is displayed because it is the number that determines what the sample can tell you. Ten items from a population of twelve is not a sample, it is nearly a census; ten from ten thousand is a sample so thin that it will find only very common problems. Neither is wrong, but they support very different sentences afterwards.
The tool does nothing about your sampling frame, and the frame is where audits actually go wrong. If the list you paste is the list of transactions that got recorded, the sample cannot say anything about the transactions that did not. Randomness applied to an incomplete list produces a perfectly fair sample of the wrong thing.
When the Random Sample From a List is fair — and when it is not
What it does guarantee
- Every subset of size n is exactly as likely as every other, which is the property a defensible sample needs.
- No item can appear twice, so the sample size is always the number you asked for.
- The sampling fraction is computed and shown, so the sample is never quoted without its denominator.
What it does not
- It cannot see what your list left out. A perfect sample of an incomplete frame is a perfect sample of the wrong population.
- It does not stratify. If your population has important subgroups, a simple random sample can miss a small one entirely.
- Sample size is your decision. The tool will happily draw three from ten thousand and say nothing about whether that means anything.
Two worked examples
12 invoices from a population of 400
- Sampling fraction is 3.0%, which the tool displays before the draw.
- The first draw is 1-in-400, the second 1-in-399, the twelfth 1-in-389.
- Every item's overall chance of inclusion is 12/400 = 3%, identical for all of them.
30 records from a population of 35
- Sampling fraction is 85.7% — this is close to a census rather than a sample.
- Only 5 records are excluded, so a defect present in 3 of 35 is almost certain to be caught.
- At this fraction the honest move is often to check all 35 and stop calling it sampling.
The jobs this is genuinely for
Internal audit and quality control are the primary case: pull a defensible random sample of transactions, records or units, and be able to describe exactly how it was drawn afterwards. The description matters as much as the draw, which is why the sampling fraction is on screen rather than left as an exercise.
Survey sampling from a member or customer list works the same way, with one extra caution: the list you have is the frame, and the people missing from it are missing from the result no matter how the draw was performed.
Teachers and trainers use it for marking moderation — pull twelve scripts at random to re-mark, rather than the twelve that were easiest to reach or that stood out. The point is precisely that the selector had no say.
If you want each item considered independently rather than a fixed sample size, the random subset picker is the correct model. If some items should be more likely than others, the weighted random picker handles that; and if you need to be able to prove the draw afterwards, run it from a published seed with the seed generator.
Frequently asked questions
What does 'without replacement' mean here?
Once an item is drawn it leaves the pool, so it cannot be drawn again. The remaining odds rise slightly with each draw, and the sample is always exactly n distinct items.
Why does the sampling fraction matter?
Because it determines what the sample can support. A 3% sample finds common problems; it will routinely miss anything affecting one item in a thousand.
Is every possible sample equally likely?
Yes. Every subset of size n has an identical chance, which is the stronger property auditors care about — not merely equal per-item chance.
Does it stratify by category?
No, this is a simple random sample over the whole list. If subgroups matter, sample each subgroup separately and combine the results.
How large should my sample be?
That depends on what you are trying to detect and is genuinely outside this tool's scope. It draws what you ask for and shows you the fraction so you can judge.
What if my sample size is larger than the population?
It is capped at the population size, which returns everything. That is a census, and the fraction will show 100%.
Can I reproduce the same sample later?
Not from this page — each draw is independent and unpredictable. For a reproducible draw you can publish and re-run, use the seed generator.
Does the order of the sample mean anything?
No. It reflects draw order only, and carries no ranking or priority.
What about items missing from my list?
They have a zero chance, and the tool cannot know they exist. That is the single most common way a sample-based conclusion turns out to be wrong.
Related tools
Random Subset Picker
Include each entry independently at probability p — the sample size is a random variable.
Random Number Generator
Any range, any count, unique or repeating — true cryptographic randomness.
Random Seed Generator
Generate a seed, then reproduce the exact same draw from it — the auditable kind of random.