How to Tell If a Political Poll Is Reliable: 8 Checks

A political poll is reliable when you can see who ran it and who paid for it, how respondents were reached, how many were interviewed and from what population, how the results were weighted, and the exact question wording. If a poll hides any of those, treat it as a press release. Even a well-made poll is one data point, so the last check is always whether it agrees with other recent polls.

Here are the eight checks we run on every poll that crosses our desk, in the order that catches problems fastest.

1. Who ran the poll, and who paid for it? #

Start with the name at the top. Established pollsters (university polling centers, major media partnerships, long-running commercial firms) put their reputation on each release. Many are members of the AAPOR Transparency Initiative, which commits them to disclosing their methods.

Then check the sponsor. A poll paid for by a campaign, party or allied super PAC isn’t automatically wrong, but it was released because the sponsor liked the result. Campaigns poll constantly and publish selectively, so a sponsored poll tells you at least as much about the sponsor’s message as about the race.

2. How were respondents reached? #

The contact method tells you who could have ended up in the sample.

Opt-in panels deserve extra caution. In a Pew Research Center experiment, 12% of opt-in respondents under 30 claimed to be licensed to operate a nuclear submarine, against roughly zero in Pew’s probability-based panel (Pew, 2024). Good opt-in pollsters screen out bogus respondents and weight carefully; some don’t. Are online opt-in polls reliable? covers the tells.

3. How big is the sample? #

Bigger is better, with diminishing returns. Approximate margins of error at 95% confidence:

Two things people miss. First, the margin on the gap between two candidates is close to double the margin on each share. A 48-45 race with a ±3.5 margin is well within noise. Second, subgroups have much smaller samples. If a poll of 800 says one candidate leads among 18-to-29-year-olds, that finding might rest on 90 people.

Weighting also inflates the real margin of error beyond the textbook number, because some respondents count more than others. Pollsters who report a “design effect” or an adjusted margin are being straight with you.

4. Who is the population: adults, registered voters or likely voters? #

This one label changes results more than most people expect.

• Adults (A): anyone 18 and over. Fine for approval ratings and issue questions, weakest for predicting an election.
• Registered voters (RV): people on the rolls. Standard until late summer of an election year.
• Likely voters (LV): registered voters the pollster expects to vote, based on a screen or turnout model. Most useful in the final two months, but every pollster builds its screen differently.

Compare like with like. A candidate who “gained 3 points” between a June RV poll and a September LV poll may not have gained anything. See registered vs. likely voters for how the screens work.

5. How was the sample weighted? #

No raw sample looks like the electorate, so pollsters weight responses to match known targets from the Census Bureau and voter files. A trustworthy poll tells you what it weighted on. At minimum you want age, gender, race and ethnicity, region, and education.

Education is the one that burned pollsters in 2016. An American Association for Public Opinion Research review of that election found that many state polls failed to weight by education, at a time when college graduates were more likely to answer surveys and increasingly voted differently from non-graduates. Many pollsters now also weight by past vote or party registration, which helps some and creates its own risks.

6. What exactly was asked, and in what order? #

Read the question as the respondent heard it. “If the election were held today, would you vote for [Candidate A] or [Candidate B]?” is neutral. “Would you be more or less likely to support Candidate A if you knew…” followed by an accusation is persuasion, not measurement. That is the classic push poll, and it isn’t a poll at all.

Order matters too. A ballot question asked after ten questions about inflation can come out differently from the same question asked first. Good releases include the full questionnaire in order.

7. When was it in the field? #

Field dates tell you what news the poll could and couldn’t reflect. A poll fielded over two weeks that ended before a major debate says nothing about the debate. Very short fields (one night) can miss people who are hard to reach. Check the dates before you react to “new” polls, because some are published a week or more after interviews ended.

8. Does it agree with other polls? #

The final test is context. By chance alone, about 1 in 20 perfectly run polls will land outside its stated margin of error. So when one poll shows a 10-point swing that no one else sees, the likeliest explanation is noise or an unusual method, not a sudden shift.

This is where a polling average earns its keep. Election Tracker, our free iPhone app, groups every published poll by race (Senate, the generic House ballot, governor, and both 2028 primary fields) and shows each one’s pollster, sample size, population and field dates on its card. The 30-day average sits on top, so you can see in a few seconds whether a headline poll is in line with the rest or sticking out.

A quick checklist you can use in 30 seconds #

1. Named pollster, named sponsor, and no campaign behind it (or you know there is).
2. Contact method disclosed.
3. Sample of at least 500 for a statewide race, 800 or more nationally.
4. Population labeled A, RV or LV, and you’re comparing it with polls of the same population.
5. Weighting variables listed, including education.
6. Full question wording available.
7. Field dates recent and listed.
8. Result roughly consistent with the average, or a clear reason it isn’t.

A poll that passes all eight is worth your attention. One that fails two or three is worth a shrug.

Frequently asked questions #

Why do two polls taken the same week show different results? #

Different populations (adults vs. likely voters), different contact methods, different weighting and plain sampling error. A 4-point difference between two good polls of the same race is ordinary. Look at the average of several polls before deciding which one is “right.”

What is the minimum sample size for a trustworthy poll? #

For a statewide race, samples under about 400 have margins of error near ±5 points on each share and ±10 on the gap, which makes them hard to use. National polls usually interview 800 to 1,500 people. Larger samples help, but method matters more than size.

Are internal campaign polls ever reliable? #

Some are well made, since campaigns pay good firms to tell them the truth privately. The problem is selection: you only see the ones a campaign chose to release. Treat a leaked internal as a signal of what the campaign wants you to believe.

How can I spot a push poll? #

It names one candidate, asks hypothetical questions loaded with negative claims, collects little demographic information and often calls huge numbers of people. Real polls ask neutral questions and publish their methods. If a “survey” seems designed to change your mind, it’s advertising.

Is a poll with a ±3 margin of error accurate to within 3 points? #

Not quite. The margin covers sampling error only, applies to each candidate’s share rather than the lead, and assumes the sample is otherwise unbiased. Non-response, weighting and turnout assumptions add error the headline number doesn’t capture.