Heading into the midterms, everyone who follows American politics had the same question: Were the polls correct, or would they underestimate Republicans once again? The answer turned out to be neither. This time around, the polls undercounted a different kind of voter: Democrats.

According to the Silver Bulletin polling averages, Democrats were on track to win the U.S. House popular vote by nine points and narrowly flip a few Senate seats in states that Donald Trump had won by double digits. Democratic voters across the country saw these numbers and braced for disappointment. They remembered the polling misses of 2016 (which underestimated Trump by about three points on average) and 2024 (which did too) and believed that they could be watching another blue mirage.

In fact, even this year’s seemingly incredible poll numbers turned out to significantly understate Democrats’ strength. Democrats are now expected to win the House popular vote by 13 points, the largest such victory in 50 years. And that four-point national polling error was small compared with the miss in certain states. In South Carolina, polls underestimated the Democratic candidate by five points on average; in Mississippi, by seven; in Texas, Georgia, and North Carolina, by nine. As a result, what was already expected to be a great night for Democrats turned into a rout of historic proportions.

Polling is hard and getting harder. Americans used to reliably answer surveys. Now simply finding people willing to pick up the phone is an uphill battle. As response rates have declined, so has pollsters’ ability to reach a representative sample of the electorate. In 2016, pollsters underestimated Trump’s support among white voters without a college degree because they couldn’t poll enough of them.

In 2020, the industry made a concerted effort to reach more of those voters. And yet the polls underestimated Trump by even more. Pollsters reasoned that even if they were now reaching the right educational groups, they must be missing certain personality types that were disproportionately likely to prefer Trump. More judgment calls needed to be made. Ahead of 2024, pollsters made a point of ensuring that their samples included enough 2020 Trump voters. That didn’t do the trick either—they managed to underestimate Trump yet again.

This cycle, pollsters introduced even more bespoke solutions. In their jointly run polls, The New York Times and Siena College devised ways to discount respondents they deemed unrepresentative—rural white college-town residents, for example, might not be like other rural white voters—and started weighting respondents based on how Republican their name sounds or what car they drive. YouGov put resources into making sure that it had enough voters from not just purple neighborhoods but deep-red ones too.

It also introduced new ways to model turnout, differentiating between a “strict” definition of likely voters and a “loose” one, because “strict” likely voters might include only the most engaged ones, and not the hard-to-reach ones who disproportionately support Trump and might still vote.

Pollsters are scientists, but they are also humans. They do their best to avoid biases, and yet they have them anyway. The adjustments they made this cycle were in service of maximizing the accuracy of their surveys, yes, but they also had another goal: not getting burned again. The changes were unsubtly aimed at not overrating the Democrats’ chances. If pollsters had used these new techniques in 2024, they would have more closely predicted the final result. Nate Cohn, the chief political analyst for The New York Times, wrote over the weekend that the paper’s post-2024 weighting techniques made its polls look better for Republicans.

In the past two midterms, in 2018 and 2022, pollsters nailed the final result with little systematic error. That’s probably because the voters that are hardest for pollsters to reach tend not to show up to vote in midterm elections, when the president isn’t on the ballot. By working hard to correct what went wrong in 2024, pollsters were solving for challenges that were less likely to come up this year anyway.

Even so, in one sense, pollsters accomplished their goal: They were not burned by hidden Trump voters. Instead, the polls moved in the opposite direction, especially when it came to some minority groups. Tracking how different groups voted so soon after an election is difficult, but according to a sophisticated new model deployed this year by the Times, Hispanic voters in Texas swung toward Democrats by 42 points—10 more points than the polls predicted.

Black voters’ nationwide support for Democrats was underrated by a similar amount. According to the polls, Democrats were set to win young men in battleground states by nine points after losing them badly in 2024. In the end, Democrats seem to have won that demographic nationwide by 24 points.

In the long history of electoral polling, errors have not tended to favor one party over the other. Then, from 2016 to 2024, they did. That they no longer do is perhaps one more sign that the Trump era is ending.