One narrow question
The margin of error describes how much an estimate could vary through sampling chance under a stated design. Imagine repeating the same sampling procedure many times. A 95% confidence procedure would produce intervals that contain the population value in about 95% of those repetitions. It does not guarantee that any single poll captures every source of error.
A worked arithmetic illustration
These figures illustrate a formula; they are not research findings. For a large population, a simple random sample and a proportion near 50%, about 1,000 completed interviews give a margin of roughly plus or minus 3 percentage points at 95% confidence. About 400 give roughly plus or minus 5 points. Clustered sampling and weighting can change the precision, so the design matters as well as the total.
Subgroups have less information
A county or age band within a national poll contains fewer interviews than the full sample. Its estimate usually has a wider margin of error. A national sample’s margin should never be pasted onto every subgroup. Look for the unweighted base beside the filtered results and treat very small bases cautiously.
What the margin does not measure
Question wording, non-response, gaps in the frame and interview mode can all affect results. The usual sampling margin does not capture those effects. For samples without known selection probabilities, a conventional probability-sampling margin also needs particular care: a formula cannot by itself establish representativeness.
Read differences in context
Two polls separated by two percentage points may be consistent with the same underlying picture, especially when their uncertainty intervals overlap. Check populations, dates, methods and question wording before treating the difference as a change. Statistical uncertainty also matters when reporting tenths of a percentage point: extra digits do not create extra certainty.
Treat any poll that reports tenths of a percent without a margin of error as marketing, not measurement.