Political Polling: Public Opinion and Campaign Research

On election night, three kinds of numbers meet one result. There are the polls, which say what people told interviewers they intended to do. There are the forecasts, which turn those polls into a probability. And there are the betting markets, which put a price on the outcome. By the morning, everyone knows which of them was closest, and the arguments begin about why. Politics is the one field where measurement, persuasion and mobilisation all face a public scoreboard, on a fixed date, in front of everyone.

That scoreboard has been unusually busy. 2024 was the largest election year in history, and its post-mortems are still being written. In the United States the state polls were the most accurate in decades, and they still underestimated the same side for the third election running. Both facts are true, and a useful account of polling has to hold them together. At the same time, the institutions that used to help the public read these numbers thinned out, Europe rewrote the rules for political advertising, and language models arrived as writers, interviewers, simulated voters and forecasters.

This project is an attempt to sort out what has changed. It is sixty-nine articles, each built on verified evidence, each honest about what that evidence does and does not show. It is not a trends report, and it does not rank pollsters, pick winners or take sides. For each new method, rule or claim, it asks the same questions. What has it actually demonstrated? Under what conditions? What did it cost, and who paid? And is the confident version that reached the headlines the version the evidence supports?

Who this is for

This project is written for people who make decisions with political numbers: campaign staff and consultants of any party, pollsters and the journalists who report their work, officials and regulators who write or enforce the rules, and citizens who want to read a poll without being fooled by it. You do not need a statistics background. The articles assume you are sharp and busy, that you prefer a plain explanation to jargon, and that you would rather hear "nobody knows yet" than a claim that collapses when you check it. Most articles end with practical advice for two audiences in particular, political parties and government, because they are the people most able to use this evidence well, or badly.

How it is organised

The sixty-nine articles fall into seven chapters.

Foundations (P0) comes first. It separates a poll from a forecast from a betting price, sets out honestly how the 2024 polls performed, and establishes the finding that frames everything after it: most campaign persuasion effects are small, so the burden of proof sits with anyone claiming a big one.

Measuring Opinion (P1) covers the craft of polling under strain: fewer people answer, the ones who do differ in politically patterned ways, and pollsters have responded with new weighting, turnout models and small-area estimates, each with its own cost.

Forecasting (P2) follows the probability. Statistical models, betting markets worth billions and language models all now claim to know who will win, and this chapter tests what each one actually knows.

Voter Data (P3) is the campaign's information layer: the voter file, the scores built on it, the gulf between American and European data law, and what happened when the big platforms stopped selling political ads in Europe.

Persuasion (P4) asks what still changes minds, now that a chatbot can write a leaflet as well as an ordinary person can, and looks hard at the few approaches whose effects last.

Mobilising (P5) prices the field. Turning people out is more achievable than changing their minds, and this chapter puts numbers on what each contact is worth, from a text message to a friend at the door.

Governing (P6) closes the corpus: the new rules on political advertising, deepfakes and data, the fight over who may price an election, the polling profession's efforts to show its work, and the old question of whether polls are good for democracy at all.

How to read it

You do not need to read in order. Each article stands alone, and each chapter has a short introduction that tells you what its articles share. Every article comes with an interactive companion that lets you try the idea for yourself: set a turnout rule, remove the polls from a campaign, walk a study through an ethics screen. If you want the quickest route in, start with the chapter that matches the decision in front of you. If you want the grounding that makes everything else easier to judge, start with Foundations.

This project has two siblings. One corpus covers the behavioural science of people: how they decide, what shapes their choices, why the same message lands differently on different minds. Another covers the craft of market research: how we measure what people think and do, and how that craft is changing. This project assumes both. It rarely re-explains voter psychology or survey method from scratch, and when it borrows a tool from its siblings it says so.

The honest note

Four commitments run through every article.

First, provisionality. Every article had to pass an admission test: the decisive development had to come after 2022. That makes the material new, and new material is uncertain. The articles say what the evidence shows, say plainly what it does not yet show, and treat "nobody has measured this" as a finding worth reporting. An article whose premise did not survive the evidence was reworked or parked, never softened into something vaguer.

Second, verified sources. Every empirical claim traces to a named study, dataset, ruling or official record, checked before it was used. The articles tell you when a claim rests on strong evidence, when it rests on a single study, and when it rests on a company's account of its own work.

Third, non-partisanship by design. Every article is written so that a campaign professional of any party, in any democracy, reads it as being about method rather than side. Examples come from many countries, and the American and European systems are treated as two natural experiments, not as a norm and a deviation.

Fourth, effective and ethical together. The methods in this corpus can find out what is true, or they can manufacture a convenient answer, and they use the same textbook either way. The articles treat the honest version and the rigorous version as the same move, because a poll engineered to agree with you is worthless as intelligence. That question runs through every chapter, and the last one makes it explicit.