Polls can help explain public opinion, but a headline number rarely tells the whole story. This guide shows how to read sample size, margin of error, methodology, question wording, field dates, and trend lines so you can compare surveys carefully and avoid treating one poll as a prediction.
Overview
A poll is a snapshot of responses collected from a defined group at a particular time. It is not a direct count of what every person thinks, and it is not automatically a forecast of what voters will do. To understand a result, start with five basic questions:
- Who was surveyed? The population might be all adults, registered voters, likely voters, residents of a particular region, or another defined group.
- How were respondents selected? A probability sample and an opt-in online sample have different strengths and limitations.
- When were responses collected? Opinions may change after a debate, news event, policy announcement, or campaign development.
- What exactly was asked? Small changes in question wording or answer choices can affect results.
- How does the result compare? A single number is more useful when placed beside earlier polls using a similar population, method, and question.
Sample size matters, but it is only one part of poll quality. A larger sample can reduce random sampling uncertainty under suitable conditions, while a poorly matched sample or unclear question can still produce a misleading impression. For a broader guide to checking numbers in news coverage, see How to Check Statistics and Claims in News Reports.
What to track
When a poll appears in local, national, or international news, record the details behind the topline result. The following comparison table provides a practical starting point.
| Factor | What it tells you | What to check |
|---|---|---|
| Probability sample | Respondents are selected through a process intended to give members of the target population a known or managed chance of selection. | Ask how the sample was recruited, contacted, and adjusted. Do not assume the label alone settles quality. |
| Opt-in sample | People volunteer to participate, often through an online panel or invitation. | Look for recruitment details, weighting, screening, and the pollster's explanation of possible selection bias. |
| Registered voters | People recorded as registered to vote in the relevant jurisdiction. | Check whether the group includes everyone registered or a modeled subset likely to participate. |
| Likely voters | A modeled group judged more likely to vote based on stated intention, past behavior, interest, or other criteria. | Find out how likely-voter status was defined. Different screens can produce different results. |
| Response rate | The share of contacted or invited people who completed the survey, where that measure is available. | Check how the rate was calculated and whether nonresponse could have affected the sample. |
| Weighting | Adjustments intended to make the sample more closely resemble selected population characteristics. | Look for the variables used, such as age, region, education, or past vote. Weighting can improve balance but cannot correct every bias. |
| Question wording | The exact language, order, context, and response options presented to respondents. | Read the full question, not only the headline summary. Note whether options were rotated or whether an explanation came first. |
| Field dates | The period during which interviews or responses were collected. | Separate the date of data collection from the date the poll was published. A poll published today may reflect earlier views. |
| Margin of error | An estimate of random sampling uncertainty for a result under stated assumptions. | Record the level and subgroup to which it applies. It generally does not capture nonresponse, wording, coverage, or modeling error. |
| House effects | A recurring tendency for one pollster's results to differ from the broader pattern. | Compare a pollster with other surveys conducted at similar times. Treat apparent patterns as clues, not proof of bias. |
| Trend lines | The direction and pace of change across comparable polls. | Use several surveys and consistent populations. A single jump may reflect method, timing, or normal variation. |
Margin of error, in plain language
A reported margin of error describes a range of uncertainty associated with random sampling. If two candidates or options are separated by less than that margin, it may be premature to describe one as clearly ahead. The calculation is usually most straightforward for a probability sample and can be harder to interpret for opt-in or heavily modeled surveys. It also does not mean that the true result is guaranteed to fall inside the displayed range.
Margins of error often apply to the full sample, while smaller subgroups can have greater uncertainty. A result among younger respondents, for example, should not automatically be treated as precise simply because the overall poll has a large sample.
Use a poll-reading worksheet
Copy this checklist into a notes document or spreadsheet whenever a poll matters to your coverage:
- Pollster: Who conducted the survey?
- Population: Adults, registered voters, likely voters, residents, or another group?
- Sample: How many people responded, and how were they recruited?
- Field dates: When did data collection begin and end?
- Sponsor: Who commissioned or paid for the poll, if disclosed?
- Methodology: Telephone, online panel, mixed mode, in-person, or another method?
- Question text: What were the exact words and answer choices?
- Uncertainty: What margin of error or equivalent limitation was reported?
- Comparison baseline: Which earlier poll or average provides a fair comparison?
- Source material: Is a questionnaire, topline release, or methodology note available?
Cadence and checkpoints
Polling is easiest to misunderstand when every new release is treated as a major turning point. A simple tracking routine creates context.
- At publication: Save the poll's questionnaire, methodology, field dates, sample definition, and topline results. Record the story's headline separately from the poll's own wording.
- Weekly during an active campaign: Compare new results with surveys using a similar population and question. Look for whether several independent pollsters point in the same direction.
- Monthly or quarterly outside peak campaign periods: Review the baseline and remove false comparisons. An approval rating among all adults should not be plotted as if it were directly equivalent to an election poll among likely voters.
- After a major event: Check field dates before claiming that the event changed opinion. A survey completed before the event cannot measure its effect.
- Before publishing a trend story: Check for changes in wording, sampling frame, weighting, mode, and likely-voter model. A methodology change can interrupt an otherwise smooth-looking trend.
When comparing polls, do not rely only on a graphic or social post. Open the original release when possible and confirm what the number represents. For guidance on comparing coverage of a developing story, see How to Compare News Sources Covering the Same Developing Story.
How to interpret changes
First distinguish a change in opinion from a change in measurement. If a result moves, ask whether the respondents, wording, dates, mode, weighting, or likely-voter screen also changed. If several comparable polls move in the same direction around the same period, the evidence for a broader shift is stronger than if only one survey moves.
Next, examine the size of the change. A small movement may fall within the poll's sampling uncertainty, especially when the two results come from different pollsters. That does not make the movement meaningless, but it does call for restrained language such as “slightly higher” or “within the poll's reported uncertainty,” rather than “a decisive surge.”
Approval ratings require the same care. Track the population surveyed, the exact question, the balance of approve and disapprove responses, and the share giving no opinion. A change in approval may reflect changes among several groups rather than a uniform shift across the public.
House effects can help explain why two surveys taken at similar times disagree. One pollster may consistently produce higher or lower results for a candidate or issue because of its sample design, weighting, mode, or questionnaire. This is not automatically evidence of intentional manipulation. Compare a pollster's current result with its own historical pattern and with contemporaneous surveys before drawing a conclusion.
Poll averages and aggregations can make trends easier to see, but they are not neutral by default. Check which polls are included, how old polls are treated, whether methodological changes are adjusted for, and whether partisan or sponsor restrictions apply. An average can reduce the influence of one unusual result while still inheriting weaknesses from the surveys it includes.
When to revisit
Revisit this guide whenever a poll becomes a trending news story, a new survey changes the apparent direction of a race, or an approval rating is used to support a larger claim. For an individual tracking file, update it at least monthly during an active campaign and quarterly during quieter periods. Review it immediately after a pollster changes its questionnaire, sampling method, weighting approach, or likely-voter model.
Editors should refresh polling guidance at least twice a year and before major election seasons or referendum campaigns. That review should confirm current pollster methodology pages, election-cycle examples, aggregation practices, and links to original questionnaires. If later results or a revised methodology changes the interpretation of an earlier poll, update the article or related coverage rather than leaving the original conclusion unqualified.
For your next poll, take five minutes to complete the worksheet, find two or three genuinely comparable surveys, and write down what has changed since the previous baseline. That habit turns a dramatic-looking number into a more useful piece of evidence—and helps readers understand what the poll can show, what it cannot show, and why the distinction matters.