How many do you need to ask?

Poll a sample instead of everyone and your answer carries a margin of errorThe half-width of the confidence interval around your result: a 40% ‘yes’ with a ±3 point margin means the true figure is very likely between 37% and 43%. — a band the true figure very likely sits in. Pick how tight you want that band and how sure you want to be, and this is how many responses it takes. Run it backwards to see what a sample you already have can promise.

The precision you want
I want a margin of ±points at% confidenceHow often the interval will contain the true value across repeated samples. 95% is the survey default; raising it widens the band, so it costs more responses.,
on an answer I expect near%, drawnA proportion's spread is widest at 50% and shrinks toward 0% or 100%. Leave it at 50% and you get the worst case — the size that's safe whatever the real answer turns out to be. frompeopleThe total group you're sampling from. For a big or open-ended audience leave this at 0 — the finite-population correction only matters when your sample is a sizeable slice of the whole. (0 = a lot).
Responses needed
1,068responses

Enough to pin a result near 50% to within ±3 points, 95% of the time, sampling from a very large population.

1,068responses needed
1,068worst case (50%)The size you'd need if the split were 50/50 — the most a proportion can ever demand. Plan to this and no real answer can catch you short.
—of the population
MarginResponses
±5 pts
385
±4 pts
601
±3 ptschosen
1,068
±2 pts
2,401
±1 pts
9,604
±0.5 pts
38,415

Precision isn't linear: tightening the margin from ±2 to ±1 point doesn't cost double — it roughly quadruples the responses. The last point of precision is always the most expensive one.

Already have the responses? See the margin a sample buys →

single-proportion margin of error · worst case at 50% · see margin of error