A poll drops. The headline says one candidate "surges to a three-point lead." Somewhere near the bottom, in grey type nobody reads, sits the line: margin of error plus or minus 3 percentage points. Those two facts cancel each other out. The headline is describing a lead that the poll itself cannot confirm exists. This is the single most useful thing you can learn about reading political coverage, so here it is in plain language. The margin of error is a measure of how much a poll's number could differ from reality purely because the pollster talked to a sample of people instead of all of them. The Pew Research Center defines it as how close you can reasonably expect a survey result to fall relative to the true population value. It is not a measure of whether the pollster is any good. It is arithmetic about sample size, and nothing else. The "plus or minus 3 points" almost always comes attached to an invisible phrase: at the 95% confidence level. Pew spells out what that means — if you fielded the same survey 100 times, you would expect the result to land within 3 points of the true population value in 95 of those runs. Which quietly concedes that roughly one poll in 20 lands outside its own margin. Not because anyone cheated. Because that is how sampling works. Sample size drives the number. Pew's worked example: a sample of 1,067 people produces a margin of error of about plus or minus 3 percentage points for the full sample. Bigger samples shrink the margin, but with diminishing returns — you do not halve the margin by adding a hundred more people, you roughly halve it by quadrupling the sample. That is why almost every national poll you will ever see hovers around a thousand respondents. It is the point where accuracy stops being worth the money. * Now the part that news anchors skip. When a poll reports a race — 48 to 45 — you are not looking at one number, you are looking at the gap between two. And the margin of error on a gap is bigger than the margin on either candidate. Pew puts it at generally about twice the individual margin: a 3-point margin per candidate becomes roughly 6 points on the difference between them. The logic is straightforward. If one candidate's share happens to run high by chance, the other's is likely running low by chance, and the two errors stack in the same direction. Run that against the headline. A 3-point lead, reported with a 3-point margin of error, actually sits inside a roughly 6-point band of uncertainty on the gap. The honest translation is not "she leads by three." It is "either of them could be ahead, and we cannot tell you which." That is a much worse headline. It is also the accurate one. It gets steeper for subgroups. Every time a story says "among young voters" or "among Hispanic voters," it is slicing that thousand-person sample into a much smaller one, and the margin of error grows accordingly. Pew's own illustration: Hispanic respondents at around 15% of a sample come out to roughly 160 cases, which carries a margin of about plus or minus 8 points for a single candidate — and about plus or minus 16 points on the difference between two. A 10-point swing in that crosstab is noise wearing a costume. Political reporters build entire narratives out of exactly these numbers. The same trap catches poll-to-poll comparisons, which is where most "momentum" stories come from. Comparing two separate polls stacks two separate margins of error on top of each other. In Pew's example, a 3-point shift between one poll and the next carries a margin of error of about plus or minus 8 points. The candidate did not surge. Two different samples of a thousand strangers disagreed by an amount that is entirely ordinary. * Here is the harder truth, and it cuts against the pollsters too: the margin of error only covers sampling error. It says nothing about question wording, about who refuses to answer the phone, about whether the likely-voter model guessed right on turnout. Pew notes these other sources of error are much more difficult to quantify and are rarely reported. So the printed margin is a floor on the uncertainty, not a ceiling. It is the error you can calculate, standing in for all the error you cannot. How much does that matter? The American Association for Public Opinion Research reviewed the 2020 US general election polls and found that among polls conducted in the final two weeks, the average error on the vote margin ran 3.9 percentage points too favorable to Biden in national polls and 4.3 points in statewide presidential polls — the largest national polling error in 40 years. Every one of those polls carried a margin of error around 3 points. The misses were bigger than the disclosed uncertainty, systematically, in the same direction. That is not sampling error. That is the stuff nobody prints a number for. One more thing worth knowing, because it is where a lot of the polls you see online actually come from. AAPOR's position is that the margin of sampling error applies only to probability-based surveys, where every person has a known, non-zero chance of being sampled. It does not apply to opt-in online polls — panels of people who signed themselves up. Some of those pollsters publish a "credibility interval" instead, which sounds reassuringly similar and is not the same thing: it is the output of a Bayesian model built on the pollster's own modelling choices. AAPOR's guidance is blunt that the public should not rely on a credibility interval the way it relies on a margin of sampling error. So what do you actually do with a poll? Four questions, in order. What is the sample size, and is this the full sample or a subgroup? Is the reported gap bigger than roughly twice the stated margin of error — and if it is not, treat the race as a tie. Is this a probability sample, or an opt-in panel reporting a credibility interval? And is the story comparing two different polls, in which case the uncertainty is bigger still? None of this means polling is worthless. A well-run poll is one of the few instruments that lets a few thousand ordinary people speak louder than the handful who get quoted by name. The problem is not the pollsters, who publish their margins honestly in the methodology note nobody clicks. The problem is the layer above them — the headline writer who needs a lead, the anchor who needs a shift, the pundit who needs a narrative. The margin of error is the moment a poll admits what it does not know. The coverage is where that admission gets thrown away.