How Polls Work and Why They Sometimes Miss: A Guide (2026)

A poll is a survey of a small, deliberately chosen group of people asked the same questions, in the same way, at the same time, and then adjusted so the group’s makeup matches the wider population. That is the whole trick. Polls estimate what a much larger group would say if you could ask everybody, and they sometimes miss because the gap between the sample and the population is never zero.

Most of the disappointment people feel about polling comes from one habit: reading a poll’s number as a count. It is not a count. It is an estimate with an error bar attached, and the error bar that gets printed next to it covers only a fraction of the ways a poll can go wrong. This guide is about how polls are actually built, which parts of the process are reliable, which parts are soft, and what to do with a result once you have it in front of you.

Table of Contents

1. How Polls Work and Why They Sometimes Miss
2. How Polls Turn a Sample of People Into an Estimate
3. What Does a Poll’s Margin of Error Mean?
4. Why Polls Sometimes Miss the Mark
5. Who gets undercounted, and why it matters
6. What Do Pollsters Mean by Voters?
7. How Question Wording Can Change the Answer
8. Why Polls Get Adjusted or Weighted
9. How to Read a Poll Before You Trust It
10. Frequently Asked Questions
11. How accurate are polls and surveys?
12. What are the main problems with public opinion polls?
13. Why do polls show different results from each other?
14. Is weighting a form of cheating?
15. What is the difference between a poll, a polling average and a forecast?
16. Why do polls not publish ranges instead of point numbers?
17. Start by Checking What Kind of Poll You’re Reading

How Polls Work and Why They Sometimes Miss

Five things have to go right before a poll produces a number you can argue with: the sample has to represent the people being measured, the questions have to be asked identically for everyone, the fieldwork has to happen close enough to the moment that matters, the weighting has to fix what the sample got wrong, and the result has to be read as an estimate rather than a fact.

Any one of those can fail quietly. A poll that gets all five right will usually land within a few points of the real thing. A poll that gets three right can be off by double digits and still be published with a confident-looking percentage.

It also helps to say what a poll cannot do. It cannot tell you how intensely people feel, why they hold a view, or what nonvoters think. A poll of registered voters is not a poll of voters. A poll fielded in June is not a poll of Election Day. And a poll has never once told anyone what the results will be, only what a slice of the population said on a particular range of dates.

How Polls Turn a Sample of People Into an Estimate

The pipeline runs roughly the same way at every reputable firm, whether they are Gallup, the Pew Research Center or a smaller house you’ve never heard of.

1. Define the population. Not everyone in the country. Adults, registered voters, likely voters, adults in one state. The choice is made before a single person is contacted, and it changes what the number means.
2. Build a sample frame. The list or system that identifies who could be in the sample. Historically that was phone numbers, selected through random digit dialing. Now it is often a voter file, a panel recruited online, or some combination of both.
3. Draw the sample. Ideally at random, so every person in the population had an equal chance of being picked. Random digit dialing and probability-based panels come closest to that. An opt-in panel, where people sign up and anyone can complete several polls a week, does not.
4. Choose a mode of interview. Live phone calls, text messages, mail ballots sent to a random sample of households, or online surveys. Mode changes answers. People disclose more on a self-administered form than they will say out loud to a stranger on a phone.
5. Write the questionnaire. Every respondent gets the same words in the same order, including the people who answer in a language other than English.
6. Field it. Interviews run across a few days, usually not more than a week. Interviewers are monitored for neutrality and for cheating, and recorded calls are spot-checked.
7. Weight the results. The raw sample almost never matches the population, so responses are adjusted to match known benchmarks such as age, education, race and ethnicity, and region.
8. Publish with a topline and a methodology note. The percentages, the sample size, the sponsor, the dates and the margin of error.

A small worked example. A poll interviews 1,000 adults and finds 52 percent support a policy. Weighting nudges that to 50 percent after the sample turns out to have too many college graduates and too few renters. The final published figure is 50 percent, plus or minus 3.1 points. Nobody counted 50 percent of anybody. They asked a thousand people, adjusted for known differences, and reported what that thousand suggests about a population of millions.

What Does a Poll’s Margin of Error Mean?

The margin of error covers one thing: random sampling error. If you drew a different thousand people from the same population, how far apart could the two results plausibly land? The standard formula is simple enough to do in your head with a calculator, and it depends almost entirely on sample size and on how close the result is to 50 percent.

Sample size	Reported margin of error	Difference in two candidates’ shares	Realistic range after non-sampling error
400	plus or minus 4.9	plus or minus 6.9	roughly 7 to 12 points
600	plus or minus 4.0	plus or minus 5.7	roughly 6 to 10 points
1,000	plus or minus 3.1	plus or minus 4.4	roughly 4 to 7 points
1,500	plus or minus 2.5	plus or minus 3.5	roughly 3 to 6 points

Two details in that table catch people out. First, when a poll reports a gap between two candidates, the uncertainty on that gap is wider than the margin of error printed beside each candidate, because both estimates carry their own error.

Second, and more important, the published figure does not cover non-sampling error. It does not cover people who never picked up, people who lied, people who were never reached, or people who changed their minds three weeks later. G. Elliott Morris has crunched more than 9,000 general election polls since 1980 and found the typical poll’s real error runs about one and a half times what it claims, with a long tail beyond that. Roughly a quarter of polls fall outside their own stated margin of error, when a true 95 percent confidence level would predict about one in twenty.

The practical rule that falls out of this: treat a reported margin of error as a floor, not a ceiling, and treat any lead smaller than twice that number as a tie.

Why Polls Sometimes Miss the Mark

Here is the full list of the ways an estimate drifts away from reality, roughly in order of how much damage each one does.

• Coverage error. Some people are simply harder to reach. Households without a landline, people who moved recently, people with limited English, people in rural areas with poor signal, people who do not appear in voter files at all.
• Nonresponse. The people who were selected and did not answer, or who hung up, or who never opened the envelope. If those people differ systematically from those who did respond, the result tilts. Response rates have fallen steadily for decades.
• Unrepresentative opt-in panels. People who volunteer for surveys are not the same as people chosen at random. They are more politically engaged and more opinionated. Panel companies screen for fraud by embedding trick questions, and Pew has documented large gaps: in one December 2023 opt-in survey, 20 percent of under-30s agreed that the Holocaust is a myth, against 3 percent on its probability panel, and 24 percent of self-identified Hispanic respondents said they were licensed to drive a nuclear submarine, against 2 percent on the panel.
• Turnout modeling. A poll of likely voters depends on a statistical guess about who will show up. Guess wrong and the poll was measuring the wrong population all along.
• Late deciders. Voters who decide in the final days, and who decide differently than they told a pollster in the spring.
• Question wording and order. Small phrasing choices move numbers by several points.
• Mode effects. The same question asked online, by phone and on paper does not produce the same distribution.
• Weighting choices. Weighting to the wrong benchmark, or over-correcting for one demographic at the expense of another.
• Subgroup samples that are too small. Most pollsters do not report a subgroup result with fewer than 100 respondents, and rightly so, because the error bar on 40 people is enormous.
• Reporting errors. A number that leaves the methodology note and gets described in a headline as a lead, a trend, or a prediction.

Who gets undercounted, and why it matters

The groups that land hardest in these error sources are not random. Lower-income adults, people with less formal education, renters, people who move often, families with prepaid phone plans, and anyone who has learned that institutions asking questions rarely have anything good coming back. That last one is not measurable from a survey and probably not measurable at all, which is exactly why it matters.

If your community is not in the sample, or is in it at a rate that does not match reality, the honest response is not to assume the pollster was trying to erase you. It is to note that the published margin of error describes the sample, not your neighborhood, and that subgroup numbers for a small group are close to meaningless.

What Do Pollsters Mean by Voters?

Pollsters use four different words and readers treat them as interchangeable.

Registered voters are people on the rolls. This is the easiest population to sample because a list already exists, but it is the least predictive, since most registered voters do not vote in a given election.

Likely voters are registered voters who a model predicts will vote. The model weighs past turnout, vote history, age, party registration and stated intent. This is where most polling error enters an election forecast, because the model is a forecast about behavior and the poll is an estimate about opinion.

Actual voters are people who showed up. Nobody polls them beforehand, which is why every pre-election poll is a prediction about this group rather than a measurement of it.

Adults is what most issue polling measures, including people who are not eligible to vote. Useful for questions about services and policy, misleading for anything about elections.

A poll of adults in a state where 500,000 people recently moved between states can be answering a question about a population that no longer exists in the same shape.

How Question Wording Can Change the Answer

Wording is not a technicality. It is one of the largest single sources of measurable movement in a poll, and it is the easiest thing for a sponsor to influence without technically breaking any rule.

Pew has a clean example. Asked whether the best way to ensure peace is through military strength, 55 percent agreed. Ask the same respondents whether the best way is good diplomacy, and agreement falls to 36 percent. Same respondents, same week, opposite framing.

Other levers readers should watch for: whether undecided respondents are offered a real option or pushed off the top line; whether a leading term like “reform” or “radical” appears; whether the order of options primes one answer; whether a question embeds a premise the respondent would have rejected; and whether refusal is even offered as a response.

None of this means a pollster is lying. It means wording is a design choice with measurable consequences, and the only real protection is seeing the exact question.

Why Polls Get Adjusted or Weighted

Weighting exists because a raw sample is almost never shaped like the population. If 40 percent of your respondents have a college degree but only 30 percent of adults do, you adjust the weights until the two match. A pollster might explain it as dialing a non-college respondent up by a factor of 1.5 so that group is represented at its proper size.

What weighting fixes: known differences in age, education, race and ethnicity, gender, region and sometimes party identification.

What it cannot fix: people who were never sampled, people who declined, people who answered dishonestly, people whose opinions shifted after fieldwork, and subgroups too small to adjust for at all.

The most serious weighting failure in recent memory happened in 2016, when national polls over-weighted white college graduates relative to the actual electorate and undercounted working-class voters in key states. Polling groups rebuilt education weighting afterward, and 2020 national polls overstated the Biden margin by 3.9 points, the largest miss since 1980 according to AAPOR’s post-election task force.

Weighting is not rigging. Forcing a survey to reproduce a known election result would be rigging, and it is the one manipulation that is genuinely easy to detect, because the numbers land too precisely on a past outcome.

How to Read a Poll Before You Trust It

Ninety seconds, eight questions, in this order.

1. Who ran it, and who paid? A named pollster with a public track record beats an anonymous release every time. Campaigns commission polls constantly and release only the flattering subsets.
2. Who was the population? Adults, registered voters or likely voters. National or one state. Anything else, like “adults who are likely to vote in the general election,” needs the model explained.
3. How many people? Below 400, treat any subgroup number as unusable. Above 1,000 and you are usually past the point where more interviews improve much.
4. How were they reached? Random digit dialing and probability panels can estimate a population. Opt-in panels can describe their members. If the release does not say, assume the worst common option.
5. When were they asked? Look at the field dates. An opinion measured three months out is a memory of a campaign that no longer exists.
6. What exactly was asked? Find the topline or questionnaire. If it is not published, you cannot evaluate the result at all.
7. What was adjusted? Which benchmarks, and were they census figures or party identification? Party weighting has quietly become standard practice and is worth an eyebrow.
8. What is the error band? Double whatever margin of error is printed before you decide whether a lead is real.

One more filter worth applying: does the pollster belong to the AAPOR Transparency Initiative, which requires releasing sample size, mode, sponsor, weighting and question order for every public poll?

Frequently Asked Questions

Start by Checking What Kind of Poll You’re Reading

Before you accept or argue with a number, find three things: who was asked, when they were asked, and how they were reached. If a poll was fielded recently, has a reasonable sample size, used random selection or a well-screened probability panel, and names its sponsor and methodology, it is a decent read on what a group of people thought on a particular set of days.

Then compare it against two or three other reputable polls instead of one. Agreement across firms is worth more than precision in any single release, and it is the fastest way to separate a real signal from an outlier with a weighting problem.

And if the result does not describe your community, that is worth saying out loud rather than discarding. A poll is a window onto a population, held up briefly and imperfectly. It is not a portrait of everyone, and it was never going to be.