How to Read an Opinion Poll
Every headline number rests on choices about sampling, weighting and wording. This guide explains how to read a UK opinion poll — and what it cannot tell you.
9 min read
An opinion poll arrives in the news as a single number — a party on 40 per cent, a policy backed by two-thirds of the public, a leader’s approval down five points. That number is the end of a long chain of decisions about who was asked, how, and when. Reading a poll well means looking past the headline to those decisions, because they determine how much the figure can actually bear. This guide sets out what to check, in plain English, and where the genuine limits lie.
Who ran it, and when
Start with the source and the date. A poll is only as trustworthy as the transparency behind it: a reputable survey discloses who commissioned it, who conducted it, how many people were interviewed, and the fieldwork dates — the days on which people were actually asked. Membership of the British Polling Council is a useful signal here, because its members agree to publish those details and the underlying data tables.
The fieldwork dates matter more than they appear to. Opinion moves with events, so a poll taken before a major announcement is measuring a different moment from one taken after it. A “new” poll reported today may have been in the field a week ago.
Sample size and the margin of error
Most national polls interview between 1,000 and 2,000 people. From that sample the pollster infers the views of tens of millions — and the inference carries an unavoidable uncertainty, the margin of error. For a typical poll this is roughly plus or minus three percentage points at the headline level.
That figure should change how you read every number. A party reported at 40 per cent could really be anywhere from about 37 to 43. So when two parties are “two points apart”, they may in truth be level. And when a party “rises one point” since last week, that movement is usually within the margin of error — noise, not news. Bigger samples narrow the margin, but only slowly: beyond about 1,000 respondents, doubling the sample buys surprisingly little extra precision.
Weighting: making the sample look like Britain
No raw sample is a miniature of the country. Some groups are easier to reach than others, so pollsters weight their data — adjusting responses so the sample matches the known population profile by age, sex, region, past vote and more. If young people are under-represented in the raw responses, the young people who did take part are given proportionally more weight.
Weighting is standard and necessary, but it is also where much of the difference between pollsters comes from. Two firms surveying the same week can produce different figures simply because they weight by different variables, or model past vote differently. None of this is manipulation; it is judgement — and judgement is exactly what a single headline number hides.
The question behind the number
Always try to read the exact question asked, not the reported summary. Wording shapes answers: a proposal described in one set of terms can draw very different support from the same proposal described in another. The answer options matter too — whether respondents were offered a middle option, or a “don’t know”, changes the distribution. So does question order, since earlier questions can prime later ones.
This is why two polls on apparently the same issue can diverge. They may not have asked the same thing.
Online, phone, and house effects
Polls are conducted by different modes — large online panels, telephone calls, sometimes face to face — and each reaches people slightly differently and can elicit slightly different answers. Combined with each firm’s weighting choices, this produces stable, recognisable house effects: one pollster tends to show a party a touch higher, another a touch lower, consistently.
The practical lesson is to distrust any single poll in isolation and to watch the polling average and the trend instead. A figure that looks dramatic on its own often sits comfortably within the range the other firms are showing.
What a poll cannot do
Even a well-run poll has hard limits. It is a snapshot, not a forecast: it records opinion at the moment of fieldwork and cannot capture a late swing. It measures breadth, not intensity — a view can command majority support yet little urgency, and a quietly held majority can lose to a motivated minority. And it cannot resolve the gap between opinion and action: as the Registry’s report on the generational divide notes, a view widely held among younger voters carries electoral weight only when matched by turnout.
Election misses almost always trace back to this last point — a sample that did not reflect who actually voted. Pollsters revise their methods after a notable error, but no method removes the uncertainty altogether.
A quick checklist
Before you take a headline at face value, ask:
- Who ran the poll, and is the methodology published?
- When was the fieldwork, and what happened around then?
- How many people were sampled, and what is the margin of error?
- What exactly was the question, and what options were offered?
- Is this one poll, or does the average of recent polls agree?
A poll read this way is a genuinely useful instrument — a disciplined measurement of public opinion. Read only as a headline, it promises a precision it was never able to offer. For how opinion turns into outcomes at the ballot box, see the guide to how MPs vote in Commons divisions; for a worked example of reading the survey evidence on one question, see the report on young voters and the EU.
Questions & Answers
What is the margin of error on an opinion poll? +
The margin of error expresses the uncertainty that comes from surveying a sample rather than the whole population. For a typical poll of around 1,000–2,000 people it is roughly plus or minus three percentage points at the headline level. A party on 40 per cent might therefore really be anywhere from about 37 to 43 per cent — which is why small movements between polls are often noise, not change.
How big does a poll sample need to be? +
Most national voting-intention polls sample between 1,000 and 2,000 people. Beyond about 1,000, increasing the sample shrinks the margin of error only slowly, so a much larger sample is not automatically a much better poll. What matters more than raw size is whether the sample is representative of the population once it has been weighted.
What does it mean to 'weight' a poll? +
Weighting adjusts the raw responses so the sample matches the known profile of the population — by age, sex, region, past vote and other factors. If a sample contains too few young people, for example, the young respondents it does have are given more weight. Weighting is standard and necessary, but the choices behind it are also a source of difference between pollsters.
Why do different polls show different numbers at the same time? +
Pollsters make different methodological choices — how they sample, how they weight, how they handle people who say 'don't know', and whether they interview online or by phone. These produce consistent small differences known as 'house effects'. The trend across many polls is more reliable than any single firm's figure.
Should I trust a single poll? +
No single poll should be read as definitive. Any one result can fall at the edge of its margin of error, and an unusual finding may simply be statistical noise. The sounder approach is to follow the average of recent polls and the direction of travel over time, rather than reacting to one striking number.
Does the wording of a question affect the result? +
Yes, significantly. The phrasing of a question, the answer options offered, and even the order in which questions are asked can all move the result. This is why it is worth reading the exact question rather than the headline summary, and why two polls on the 'same' issue can diverge if they asked it differently.
What is the British Polling Council? +
The British Polling Council is the body whose member firms agree to publish the methodological details behind their polls — sample size, fieldwork dates, the questions asked and the data tables. Membership is a useful first check: it means the information needed to judge a poll is disclosed rather than hidden.
Can a poll predict an election result? +
A poll is a snapshot of opinion at the time of fieldwork, not a forecast. It cannot capture late swings, differential turnout, or how undecided voters will break. Polls describe where opinion stands; turning that into a prediction requires further assumptions, each of which can be wrong.
Why do polls sometimes get elections wrong? +
Errors usually come from the sample not reflecting who actually votes — most often because turnout differs from what was assumed, or because certain groups are systematically under- or over-represented even after weighting. Pollsters revise their methods after notable misses, but no method removes uncertainty entirely.
What can a poll not tell you? +
A poll cannot tell you how strongly people hold a view, whether they will act on it, or what they will think next month. A position can command majority support in the abstract yet command little intensity, low priority, or low turnout — all of which matter as much as the headline percentage.