Every Poll Is a Model Now

I am frequently asked my opinion on the polls in the news. It comes with the territory. AlphaVu has spent 15+ years analyzing and predicting public opinion, and election season makes the question constant.

It usually sounds like this: "Are the polls right?" My answer starts somewhere else. Which poll? And what did the pollster assume?

That's not a dodge. It's the single most important thing to understand about polling today. In this two-part series, I want to explain why, and leave you with a simple framework for reading the polls you see in the news and in your feed. Part 1 covers how polls actually work now. Part 2 covers how to read them without getting played.

How a modern poll works

Every poll starts with a population: the group you want to understand. That might be all customers, adults, registered voters or likely voters. Those are different groups, and they can give you different answers.

Pollsters can't talk to everyone, so they draw a sample. The goal is a representative sample, one where every member of the population has a known chance of being selected, so a small group mirrors the whole. The old pollster's analogy still holds. You don't have to eat the whole pot of soup to know how it tastes. One spoonful will do, as long as the pot is stirred.

Then comes the real work:

• Reaching people. Live phone calls, text messages with a link to an online survey, online panels, even mail.
• Weighting. No sample comes back perfect. Pollsters adjust the results so the sample matches the population on known traits: age, gender, race, education, region and, more and more, how people voted last time.
• Modeling the electorate. Election polls also have to estimate who will actually vote. That likely-voter model is often the single biggest assumption in the poll.

The math didn't change. We did.

The statistics behind a good poll are the same as they were a generation ago. What changed is us.

In 1997, Pew Research Center got a response from 36% of the households it called. By 2018, it was 6%. Today, at NYT/Siena, one of the most accurate polls in the country, fewer than 2% of the people its callers try to reach will respond. One of its 2022 phone surveys reported a response rate of 0.4%.

Sources: Pew Research Center (2012, 2017, 2019 reports); NYT/Siena 2022 figure as cited in Pew's Sept. 2026 weighting report.

Two forces drove that collapse. The first is trust. Americans have grown far more skeptical of institutions, of the media, of pollsters and of anyone calling from a number they don't recognize.

The second is technology. Robocalls hit an estimated 3.4 billion a month by 2018. Phones now flag or block unknown numbers automatically, and poll calls get caught in the net.

Here's why that matters. A small sample isn't the problem. Statistics handles small samples just fine.

The problem is who answers. The people who pick up tend to be joiners, more engaged in civic and political life than the people who don't. When the people willing to take a poll think differently from the people who won't, the poll is skewed before the first question is asked.

Pollsters call this nonresponse bias. It's the central challenge of modern polling.

Every poll is a mini-model

When only a sliver of people respond, a pollster can't just count the answers and report them. They have to correct for who's missing. That means building assumptions directly into the poll: what the electorate will look like, which traits to weight on, how many Democrats, Republicans and independents belong in the sample.

So every modern poll is really a mini-model.

That isn't a knock on polling. It's the reason good polling still works. Almost 70% of national polls in 2024 weighted on party affiliation or past vote, and Pew finds the practice has made polls modestly more accurate. When Pew applied it to its own postelection polls in 2020 and 2024, accuracy on the margin between candidates improved by about 2 to 3 points.

Those assumptions aren't gut feelings. The best pollsters build them from deep experience, years of their own data and hard lessons from past misses. That judgment is now what separates great pollsters from mediocre ones.

Here's evidence it works. In 2022, The New York Times ran an experiment in Wisconsin. It mailed voters a survey and paid up to $25 to respond. Nearly 30% did, compared with the roughly 2% who answer its phone polls. The two groups looked similar on how they planned to vote.

A low response rate doesn't make a poll wrong. A bad model does.

It also explains something that confuses a lot of people. Two good pollsters can look at the same race and publish different numbers. Neither is necessarily wrong. They made different assumptions, and that spread is information, not failure.

What the margin of error does and doesn't tell you

Every poll reports a margin of error. Most people read it wrong.

• It applies to each candidate, not the lead. In a 48–45 race with a ±3 margin, each number can move by three.
• It only covers sampling error. It doesn't capture a bad turnout model, a slanted question or a group the sample missed. A study of 4,221 polls across 608 state-level races from 1998 to 2014 found real-world error ran about twice the reported margin.
• Outliers are built in. At the standard 95% confidence level, about one poll in 20 will land outside its own margin, even when everything is done right.

Keep one more distinction straight. A poll is a snapshot of opinion on the days it was taken. A polling average combines many snapshots. A forecast models all of it to estimate an outcome.

When a forecast gives a candidate a 70% chance, it's saying they lose three times out of ten. That's not a lock.

From mini-model to digital twin

The same idea sits at the foundation of AlphaVu Predict, taken a step further.

A modern poll pairs survey data with a model of the audience. Predict does the same thing, continuously. It builds a digital twin of an audience (customers, riders, residents, voters) anchored in survey research and fed by real-time opinion data. Then it predicts how that audience will respond to a decision before the decision goes public.

The difference is what happens after the survey closes. A poll is frozen the day it's fielded. Predict tests its own predictions against real opinion as it comes in, fine-tunes itself when accuracy drifts and signals when it's time for a new survey (patent pending). That's survey-grade confidence without surveying every week.

Same discipline. Same rigor about assumptions. Applied to every decision, not just Election Day.

Next: how to read polls without getting played

If every poll is a model, the useful question isn't "Is this poll right?" It's "Are these assumptions any good?"

Part 2 shows you how to answer it: why there's so much noise around polls, which pollsters have earned your trust, why "the polls are biased" is usually the wrong conclusion, and five questions to ask before you share a poll.

Don't guess in public. Predict first.

Scott Wilkinson is Founder and CEO of AlphaVu. Originally published on LinkedIn.

See how AlphaVu Predict works for audiences like yours →