The polling average won. Now what?

The polling average won. Twenty years ago, in The Washington Post, I worried that the new fashion for averaging treated every survey as an equal ingredient, and I wrote, “ We need standards, not averages .” Today the averages update by the hour, prediction-market odds move continuously, and a Senate margin can change before lunch. We have a better sense than ever of who is up and who is down, but a diminished sense of why.

We are not better informed about the campaigns. Seventeen percent of Maine voters who disapprove of President Trump plan to vote for Sen. Susan Collins, the Republican incumbent. In a race this close, they could decide who controls the Senate. In 2020, Joe Biden carried Maine by 9 points, and Collins still won by nearly as much. The polls tell us these voters are there. The averages and markets leave their reasons out of the picture. That should be the beginning of an inquiry.

A September New York Times/Portland Press Herald/Siena poll had Collins up 3 points, inside the margin of error. Most polls since show the race tied or give Democrat Troy Jackson an edge. The split among her own voters has held: About a third of Collins supporters disapproved of Trump, and only half said they’d rather have a senator who supports him. The tables offer clues: Voters saw her as the more effective senator, and nearly two-thirds of her voters called Jackson too far to the left. In an early-September CNN poll , a majority of Collins voters said their vote wasn’t a message about Trump at all. But no published table shows how Collins voters who disapprove of Trump answered any of these questions. The coverage stops at the margin.

On Nov. 3, some of the estimates will be right, and some will be wrong; we will spend weeks arguing about which. That’s the wrong fight. The real battle is over what the results mean. Who has a mandate, and for what? What kinds of candidates do voters want, and which positions? That debate usually runs on anecdote and spin. It should run on data, and the data exist.

Take data centers. No single poll tells that story. A dozen do. Gallup , Fox News , Emerson , and The Economist/YouGov all found national majorities this year against a data center near home, from 63 to 71 percent; Quinnipiac found 74 percent in Pennsylvania. Heatmap and Embold Research have asked the same question four times since August 2025, and opposition rose from 42 percent to 75 percent. A candidate who wins on that could claim a mandate to stop data centers from being built in their constituents’ backyards. Read the rest of the tables, and the mandate looks different. The common thread is the electric bill.

Pew found Americans who had heard of data centers six times as likely to call them bad for home energy costs as good. Six in 10 told YouGov a nearby one would raise what they pay. And when Echelon Insights asked voters to assume data centers would cover their own grid costs so bills did not rise, 54 percent said local officials should allow them; 34 percent said ban them. Opposition is a headline. Together, these polls suggest a different question for the winners: Would voters accept data centers if developers paid the costs?

Most of those tables are public. AI can search them in seconds. It can bring findings from a dozen polls together and help us see where the answers converge, where they differ, and what no one has asked. AI can also misread a column or mix up two groups. That CNN poll found that 30 percent of Collins voters disapproved of Trump. Turn the question around, and 20 percent of Trump disapprovers supported Collins. Both are true. Confuse the denominators, and you will tell the wrong story.

So the rules I asked for in 2006 still apply: disclosure from pollsters, judgment from newsrooms. When AI turns up an interesting number, a reporter should be able to see the exact question, who answered it, and where it appears in the tables. If any of that is missing, the number isn’t ready for print. If it checks out, use it to write the next question. Recontact willing respondents, or draw a fresh sample with enough of these cross-pressured voters to hear them out. Let them explain, in their own words, why they vote the way they do and what could change their minds. A campaign teaches us things. Our questionnaires should learn, too.

I was too neat in 2006. Good averages have earned their place. But they were never built to tell us what voters want, and the markets price outcomes, not reasons. Both have won. The next step is to dig beneath the surface. When the tables are public and the tools are this good, stopping at the margin is a lazy choice.

Jon Cohen founded TrueDot, an AI-accelerated research company. He was The Washington Post’s polling director from 2006 to 2013 and later was vice president of research at the Pew Research Center and chief research officer at SurveyMonkey.