Poll herding is when pollsters, consciously or not, adjust their methods or hold back results so their numbers land close to what other polls are showing. The effect is a field of polls that agree with each other more than random sampling allows. That looks reassuring, but it hides real uncertainty and can make every poll wrong in the same direction at once.
Herding tends to show up in the final weeks of high-profile races, which is exactly when most people start paying attention. Here’s why it happens, how to recognize it, and why a poll that breaks from the pack isn’t automatically wrong.
Why should honest polls disagree? #
Every poll interviews a sample, not the whole electorate, so every poll carries sampling error. With 800 respondents, the margin of error on each candidate’s share is about plus or minus 3.5 points at the usual 95% confidence level.
The margin between two candidates bounces around roughly twice as much as either candidate’s share. So in a race that’s truly tied, a set of honest polls with 800 respondents each should produce margins scattered across several points in both directions, with an occasional poll showing a clear lead for one side purely by chance. Our explainer on margin of error in political polls walks through the arithmetic.
When a dozen polls from different firms, using different methods, all land within a point of each other, that tight cluster is itself a warning sign. Random chance rarely produces that much agreement.
Why do pollsters herd? #
Nobody needs to coordinate for herding to happen. The incentives do the work.
Being alone and wrong is expensive #
Picture a pollster whose raw data shows a seven-point lead in a state where every other recent poll shows a tie. If they publish and turn out wrong, they alone take the blame, and media partners and clients notice. If they publish and turn out right, they still spend weeks defending an “outlier.” If they nudge their numbers toward the consensus and everyone misses together, it becomes an industry-wide miss that no single firm owns.
Weighting gives room to nudge #
Raw survey responses always get weighted to match the expected electorate by age, education, race, region and sometimes party or past vote. Pollsters also have to decide who counts as a likely voter. Each choice is defensible on its own, and together they can easily move a margin by several points. A pollster who tries a few reasonable options and picks the one closest to the average has herded without ever changing a single answer.
Some polls just don’t get published #
The quietest form of herding is the file drawer. A pollster, or a client who commissioned the poll, looks at a strange result and decides not to release it. The public never sees the poll that broke from the pack.
What does herding look like in practice? #
The 2024 presidential race is the recent example. In the final weeks, polls in the seven swing states clustered unusually close to a tie, and several analysts pointed out that there were fewer outliers than chance would predict. Donald Trump then won all seven, by margins from about 0.9 points in Wisconsin to about 5.5 in Arizona. Most of those misses were within normal polling error, but they all ran the same way, which is what correlated error from herding would produce.
Independence cuts both ways, though, and Iowa pollster Ann Selzer shows it. Her final 2020 Iowa poll showed Trump up 7 when others showed a closer race, and he won by about 8. Her final 2024 poll showed Kamala Harris up 3; Trump won Iowa by about 13, the largest error of her career. Publishing an outlier doesn’t make a poll right. It makes the field of polls more honest about how uncertain things are.
How does herding distort the picture? #
False precision. A tight cluster of polls makes a race look more knowable than it is. If the true uncertainty is several points and the polls imply one, readers and campaigns are overconfident.
Hidden movement. A pollster who picks up a real late shift may smooth it away to match older polls, so the shift only becomes visible on election night.
Errors that stack. Averaging is powerful because independent errors cancel out. If polls are herding, their errors aren’t independent, and an average of them can’t cancel anything. It just repeats the shared mistake more confidently. That’s why how polling averages work depends so much on the polls going into them.
How to spot herding and read polls smarter #
You can’t see a pollster’s unpublished data, but you can look for patterns.
- Distrust perfect agreement. Many polls using different modes (live phone, text-to-web, online panels) that all show the same margin in a volatile race is a sign to widen your own sense of uncertainty.
- Check field dates. Polls released in the same week often have overlapping field periods, so they aren’t as many separate looks as they seem.
- Watch individual pollsters over time. A firm whose numbers jump toward the average right before the election, after weeks of showing something different, deserves a second look.
- Value transparent outliers. A pollster who publishes a surprising result along with full crosstabs is giving you information, even if the result turns out wrong. Our guide to reading poll crosstabs helps here.
- Know each pollster’s lean. Consistent differences between pollsters are a separate issue, covered in what a pollster house effect is.
Election Tracker is built around several of these checks. Every poll card shows the pollster, sample size, population and field dates, so you can see at a glance whether a burst of agreement came from one week of fieldwork. Its 30-day average collapses repeated polls from the same pollster into one, at the strongest sample, so a firm that releases a poll every few days can’t drown out everyone else.
The app also has a Market Sentiment tab with current Polymarket prices for races such as Senate and House control. Markets have different incentives from pollsters, so comparing the two can show when they disagree. The app labels those prices as trader sentiment rather than forecasts, and traders read the same polls, so markets herd too.
Frequently asked questions #
Is herding the same as polling bias? #
No. Bias is a systematic error that favors one side, often from reaching the wrong mix of voters. Herding is the suppression of normal variation between polls. A herded set of polls can be biased or unbiased; its main harm is hiding uncertainty and making many polls miss together.
Is herding a form of fraud? #
Usually not. Most herding comes from defensible methodological choices, like picking one likely-voter model over another, made with an eye on what other polls show. Fabricating data is a different and much rarer problem.
When is herding most likely? #
In the last few weeks of high-profile races with lots of public polling, such as presidential swing states and marquee Senate races. Low-profile races have too few polls to herd toward, so their results tend to scatter more naturally.
Does using a polling average fix herding? #
Only partly. An average smooths out independent random error, but if the polls are herded, their errors are shared, and the average inherits them. The average is still the best single number, but you should treat its implied precision with some skepticism when the underlying polls agree suspiciously well.
How should I read a race when every poll agrees? #
Treat it as closer to “we’re not sure” than to certainty. If the average shows a one-point race, a result a few points either way shouldn’t surprise you. Look at the full range of recent polls and remember that a tie in the polls isn’t the same as a tie in the electorate.