How to read a poll critically

Polls drive a huge share of political news coverage and are routinely mis-interpreted. A two-minute checklist catches most of the misuse.

1. Who paid for it?

The first thing to find in any poll report is the sponsor. A poll commissioned by a campaign about its own race is not the same kind of evidence as a poll by a neutral academic institution. The reputable pollsters (Pew, Gallup, AP-NORC, Marist, Quinnipiac, Monmouth) get rated by 538/ABC’s pollster ratings — that’s a quick credibility shortcut.

2. What was the sample?

Two things matter: size (n) and type. A poll of 800 likely voters is meaningful for a US national race. A poll of 800 *registered* voters in March is meaningful for March but means less for November turnout. A poll of 200 anyone-with-a-pulse is statistically thin.

Watch for: “adults” vs “registered voters” vs “likely voters” — these are different populations and produce different results. The right framing depends on what question is being asked.

3. What was the margin of error?

Almost every poll has a margin of error around ±3–4 percentage points. A poll showing a 49-47 race has the two candidates statistically *tied*, not the 49-camp ahead. News headlines that ignore the margin and write “X leads” on a within-margin poll are doing political work, not reporting.

4. How was the question worded?

Question order matters. Question wording matters more. “Do you support a tax cut?” gets different results than “Do you support a tax cut that would primarily benefit high earners?” Both can be defensible questions; they measure different things. Reputable pollsters publish their full questionnaires — read them.

5. How was the poll conducted?

Phone, online panel, mixed-mode, opt-in vs probability sample — each has tradeoffs. Probability-sample phone polling remains the gold standard for unbiased sampling but is expensive and slow. Most modern polling is mixed-mode for cost reasons; this is fine, but pollsters who lean too heavily on opt-in online panels for political polling produce systematically less reliable numbers.

What partisan coverage gets wrong

The most common abuse is outlier cherry-picking — citing the one poll that supports a preferred narrative while ignoring the dozen others showing a different picture. Default to averages (RealClearPolitics, 538, The Race to the WH) rather than any single poll. The average is almost always closer to the eventual outcome than any individual poll.

Prism Source Receipts will verify whether a polling claim cited in an article matches the underlying poll’s actual numbers — catches the most common misuse in under 10 seconds.

Related: Primary source · False balance · How to fact-check a news article.