A poll is an estimate, not a count
An opinion poll asks a relatively small group of people a set of questions and uses their answers to describe a much larger group. The large group might be all adults in a country, registered voters, or people who use a particular service. The small group is the sample. Everything interesting about polling comes from the gap between the two: how the sample was chosen, who actually answered, and how the answers were turned into a published figure.
Because of that gap, a poll result is best read as an estimate with some uncertainty around it. Two polls can both be competently run and still land in different places, and that is often not a sign that either one is broken.
The main design choices
Every poll involves a series of decisions. Each one can shift the result a little, and together they explain most of the differences you see between pollsters.
Who is the target population?
A poll of “adults” and a poll of “likely voters” are asking about different groups. Even within one label, definitions vary. Some polls include only people who say they are certain to vote, others use a looser screen. When a headline compares two polls, it is worth checking whether they were describing the same people at all.
How is the sample drawn?
In a probability sample, every person in the target population has a known, non-zero chance of being selected, usually through a list or a randomised method. This is the traditional foundation for the statistical tools used to describe uncertainty. In practice, the list used to draw the sample is rarely perfect. The gap between the list and the real population is the subject of the discussion of sampling frames, and it matters because people missing from the list have no chance of being asked.
Many modern polls instead use volunteers who joined an online panel. These are cheaper and faster, but the people who join differ in various ways from those who do not. How survey panels work and who they leave out is a topic of its own, and the contrast with randomly drawn samples is covered in the article on opt-in polls in this cluster.
How are people contacted?
Telephone, post, face-to-face visits, text invitations and web links all reach different kinds of people. Some groups are easier to reach by phone, some by post, some online. Interviewers can also influence answers on sensitive topics, whereas self-completed web surveys tend to feel more private. Pollsters often combine methods to widen who they reach.
What exactly is asked?
Small changes in wording, order and answer options can move results. Offering a “don’t know” option, placing a question after a related one, or describing a policy in different terms may all produce different distributions of answers. The effect of question wording on survey answers is large enough that comparing polls without comparing their questions is risky.
From raw answers to a published number
Who actually answers
Not everyone who is selected takes part, and the people who decline are often different from those who respond. This is non-response, and it is one of the central concerns in modern polling because response rates for many survey types have typically fallen over the years. If the missing people differ systematically from the respondents, the result can lean in a particular direction.
Weighting
Pollsters usually adjust the raw responses so that the sample better matches known features of the population, such as age, sex, region or education. This is survey weighting. It can correct obvious imbalances, but it relies on the pollster choosing the right characteristics to adjust for and on reliable benchmarks to adjust towards. Different but reasonable choices produce different published figures from the same raw data.
Sampling uncertainty
Even a perfectly drawn random sample would differ from the true population value by chance. Pollsters often report this as a margin of error. It describes only random sampling variation, and it typically does not capture problems such as bad wording, non-response or a flawed list. A small difference between two polls is therefore often within the range you would expect from chance alone.
Why two careful polls can disagree
When polls differ, the cause is usually one or more of the following:
- Different populations. Adults, registered voters and likely voters are not the same group.
- Different samples. Random-digit or address-based sampling versus opt-in panels can attract different kinds of people.
- Different wording. Even slight changes can shift how people respond.
- Different timing. Opinions can move, and events between fieldwork dates may matter.
- Different weighting. Each pollster makes its own decisions about what to adjust for.
- Chance. Random sampling variation alone produces some spread.
None of these implies bad faith. They are the ordinary consequences of making many reasonable decisions in a setting where there is no single correct recipe.
Honest answers are not guaranteed
Respondents do not always report what they truly think or do. People may want to appear consistent, informed or socially acceptable, and they may misremember what they did. That tension between stated and actual behaviour is discussed more broadly under self-reported behaviour, and the pressure to give acceptable answers has its own article in this cluster.
Reading a poll sensibly
A few habits help when you meet a new poll result:
- Look for who was surveyed, how, and when.
- Check the exact question wording and the answer options.
- Treat small gaps as possibly within normal sampling variation.
- Prefer patterns across several polls over a single dramatic result.
- Be more cautious when a poll was a self-selected web vote rather than a designed sample.
Polls are useful tools for describing broad opinion, but they rarely give a single sharp answer. Trend stories that lean on one poll often lose these caveats along the way, which is part of why trend pieces can overstate what the data shows. A cautious reader treats each figure as one well-informed estimate among several, with its own strengths and limitations.