If you’ve been following the U.S. Senate runoff between Barry Moore and Jared Hudson, you’ve felt the whiplash. Four public polls in roughly two weeks. Four different answers.

The Alabama Poll surveyed 600 likely runoff voters on May 28 and found Hudson ahead 48.7% to 39.2%, with 12.1% firmly undecided. Days earlier, Remington Research, polling 722 voters for a pro-Hudson PAC, called it a one-point race, 41% to 40%. Co/efficient, polling June 3–4, found Moore up eight, 43% to 35%, with 22% undecided. And Strategy Management’s independent survey of 1,300 likely runoff voters, conducted June 5–7, has Hudson up nearly five, 42.2% to 37.5%, with one in five undecided.

That’s a 17-point spread between the extremes, and it invites a fair question: how can surveys of the same race, taken days apart, disagree this much? Is somebody just wrong?

Not necessarily. The more likely explanation is that each pollster is answering a slightly different question, because polling is a science wrapped around a series of judgment calls.

The science is the part everyone knows. Draw a random sample of 600 voters and the laws of probability say each number carries a margin of error of about four points, meaning the gap between two candidates can swing roughly twice that on sampling error alone. Two honest polls of the same race can differ by several points and both be “right.”

The art is everything else, and the art is where these four polls part company.

Start with the hardest call in the business: who will actually vote? Runoffs are the lowest-turnout elections we hold. Most of the Alabamians who cast a ballot on May 19 will stay home on June 16, and every pollster has to decide which ones return. Here’s the wrinkle most readers never see: every one of these polls screened for likely voters, meaning every single respondent said they plan to vote. Where pollsters part ways is in how much they trust those answers, and whose. Some of us anchor on demonstrated behavior — did this voter actually show up for past primaries? Others weight their sample to a forecast of what the electorate will look like, deciding in advance, for example, what share of the vote seniors will cast. Neither approach is dishonest. They are different bets. Two polls can interview the same state, find the same race, and publish different toplines because they presume different electorates. That is the art in this process, and reasonable professionals land in different places.

Second, the undecideds. One poll says 12% haven’t made up their minds; another says 22%. Some pollsters push leaners toward a choice; others take “undecided” at face value. The same voter who tells one interviewer she’s undecided gets nudged into a column by another. Different toplines, identical electorate.

Third, timing. A poll is a photograph, not a film. A survey fielded May 28 and a survey fielded June 7 captured different races; in between came new endorsements, heavy ad spending, and a string of headlines. In the closing days of a runoff, ten days is a lifetime.

Fourth, the mode. Live telephone interviews, text-to-web, automated calls: each method reaches a different slice of the electorate, and each has well-documented tradeoffs that professionals spend careers managing. But the mode debate gets more credit than it deserves. Strategy Management collected its interviews through both live calls and text-to-web, and both methods found the same leader. The Alabama Poll’s May survey likewise blended phone and text interviews, and they agreed too. The test of any poll isn’t how the interviews were collected — it’s whether the final sample looks like the electorate that actually shows up. Modeling an Alabama runoff electorate is learned the hard way, cycle after cycle, and that local mileage matters more than the technology.

Now, here’s the part getting skipped past. Look beneath the dueling toplines and these polls agree far more than they disagree. Every survey of this runoff shows the same two coalitions: Hudson running strongest with voters under 65, in the Birmingham market, and with independent-minded primary voters; Moore running strongest with seniors, self-described Trump Republicans, and his home Wiregrass. The polls don’t disagree about who supports each candidate. They disagree about how many of each group will show up on June 16. That’s the whole ballgame.

So how should a reader handle the next poll that crosses the feed? Ask a few questions. When was it in the field? How many voters, with what margin of error? How were likely voters screened, and how were undecideds handled? Who paid for it, and does the release say? A poll that answers those questions plainly deserves more weight than one that doesn’t, whatever its topline says.

I’ll add one more, in the interest of full candor: what’s the pollster’s track record? No one in this business gets every race right, and I’ve made my share of wrong calls over thirty years around Alabama politics. Humility is a job requirement. But disagreement among polls isn’t evidence that polling is broken. It’s evidence that reasonable professionals made different judgment calls about an electorate nobody has seen yet.

On June 16, we get the only poll with no margin of error.