The questionnaire in research: from the 1936 Literary Digest fiasco to the Likert scale, NPS, and the post-response era of 2026

A questionnaire is a research instrument consisting of a structured set of questions aimed at obtaining information from a group of people in a systematic way. It is probably the oldest and most widely used tool in social research, marketing, public health, business management, and political science. Its apparent simplicity hides a rigorous discipline: there is a century of theory on how to design good questions, how to select samples, how to avoid biases, and how to interpret results that seem clear but may be capturing noise.

This article covers the field's intellectual history, the fundamental concepts that anyone who designs questionnaires should know, and the specific challenges of the present moment, where response rates have collapsed and the quality of many commercial surveys is debatable.

The birth of the scientific survey: 1935-1936

The history of the modern survey as a rigorous discipline has a founding case that remains the canonical example taught in any methodology course.

In the U.S. presidential election of 1936, two organizations tried to predict the outcome: the magazine Literary Digest and the newly created American Institute of Public Opinion of George Gallup.

Literary Digest sent 10 million survey ballots to a list built from telephone directories, automobile owner records, and magazine subscribers. They received approximately 2.4 million responses. With that massive sample —enormous by the standards of the time, and unthinkable today— they predicted that Alf Landon would beat Franklin D. Roosevelt with 57% of the vote.

George Gallup, by contrast, had founded the American Institute of Public Opinion in 1935 with a different methodological approach: instead of giant samples from biased sources, he used stratified random sampling with just 50,000 respondents representative of the voting population. He predicted FDR would win with 56%.

Roosevelt won with 60.8%. Gallup had been right with a sample 200 times smaller than Literary Digest's, because his sample was representative while Literary Digest's was biased toward voters with a phone and a car —demographics that, in the middle of the Great Depression, over-represented the more prosperous, who tended to vote Republican.

The fiasco sank Literary Digest, which closed in 1938. And it enshrined the founding principle of the scientific survey: sample size matters less than representativeness. A survey with a small but well-selected sample predicts better than a gigantic but biased one.

This case is taught because it illustrates a mistake that is still made in 2026 in countless studies and commercial surveys: confusing volume with representativeness. A survey of your email list generates a large sample, but not necessarily one representative of your target market.

The Likert scale: 1932

Even before the Literary Digest fiasco, in 1932, social psychologist Rensis Likert published in Archives of Psychology his doctoral thesis titled "A Technique for the Measurement of Attitudes." In that work he presented what is now universally known as the Likert scale —the tool that makes it possible to measure attitudes and opinions with a graded response system.

The classic Likert scale has five points:

1. Strongly disagree
2. Disagree
3. Neither agree nor disagree
4. Agree
5. Strongly agree

Likert found empirically that this gradation captured far more nuanced information than binary (yes/no) questions and allowed sophisticated statistical analysis of attitudes.

Later variants have used 7-point scales (more sensitivity), 4-point scales with no midpoint (forcing a position, avoiding the neutral response as an evasion), and 10-point scales (especially for customer satisfaction). Each variant has arguments in its favor; the academic consensus is that consistency within a single questionnaire is more important than the exact number of points.

The Likert scale remains, almost a century later, the dominant format for questions about attitudes, opinions, satisfaction, and agreement in serious questionnaires.

NPS: Fred Reichheld, HBR, 2003

The 20th century ended and the 21st introduced another important methodological contribution. In December 2003, Fred Reichheld published in Harvard Business Review the article "The One Number You Need to Grow," in which he proposed that a single question could predict a company's growth better than any complex battery of satisfaction measures.

The question is:

"On a scale of 0 to 10, how likely are you to recommend [company/product] to a friend or colleague?"

Responses are classified:

• 9-10: Promoters (loyal customers, willing to recommend).
• 7-8: Passives (satisfied but not enthusiastic).
• 0-6: Detractors (dissatisfied, a possible source of bad word of mouth).

The NPS (Net Promoter Score) is calculated:

NPS = % Promoters − % Detractors

Reichheld worked at Bain & Company and had developed the model in collaboration with the software company Satmetrix. The patented methodology caught on extremely fast: by 2010, NPS was already a standard metric at thousands of Fortune 500 companies.

The academic criticism has been considerable: later academic studies (Keiningham, Aksoy, and others) have shown that NPS does not predict growth more reliably than other satisfaction measures. But the practical usefulness of NPS lies in its simplicity: a metric easy to communicate to non-technical executives and comparable across companies and periods. That practical virtue —communicability— has kept NPS popular even in the face of methodological criticism.

For a company, deciding between NPS, CSAT (Customer Satisfaction Score), CES (Customer Effort Score), or other metrics is not trivial. The choice depends on what you want to measure: loyalty and advocacy (NPS), point-in-time satisfaction (CSAT), ease of the experience (CES).

The big problem of 2026: the collapse of response rates

There is a fundamental change in the reality of questionnaires that the classic theories did not anticipate: response rates have collapsed over the last 50 years.

Data from the Pew Research Center and other methodological studies document the decline:

1970s (telephone surveys): response rate of 80% or higher. When the interviewer called, most people answered and participated.

1990s: the rate fell to roughly 40%.

2000s-2010s: 20-30%.

2020-2026: in commercial telephone surveys, rates are usually below 10%, and in many contexts below 5%. For non-incentivized online surveys (not a paid panel), rates can be 1-3% of the initial send.

This decline has several converging causes: saturation (people receive many more survey requests), spam-call fatigue (rejecting unknown numbers by default), loss of trust in how personal data is used, and the proliferation of low-quality commercial surveys that have devalued the category.

The consequences for honest research are significant:

Amplified non-response bias. If only 5% respond, the characteristics of those who respond can be very different from those of the 95% who don't, and the sample stops being representative.

Field costs rose. Reaching a representative sample requires many more calls/sends per response obtained, significantly raising the cost of rigorous research.

Growth of paid online panels. Platforms like Prolific, CloudResearch (formerly TurkPrime), Cint, Toluna, and Lucid offer access to paid panelists who do respond. This solves the response problem but introduces a new bias: panelists are a specific subpopulation.

Growing reliance on predictive models. To compensate for low representativeness, serious firms use statistical modeling (post-stratification, Multilevel Regression with Poststratification) to correct known biases. But these models only correct biases that are anticipated.

For a company designing commercial surveys (not rigorous academic studies), the practical implications are:

• A survey of "your email base" is not representative of your market, only of your current customers.
• An open survey on social media captures whoever cares enough to respond, not the general audience.
• An incentivized survey biases toward those who value the incentive.
• Honest conclusions require explicitly acknowledging these limitations, not presenting results as if they were representative when they aren't.

Modern tools and their operational reality

The survey software market in 2026 is dominated by a few notable platforms:

SurveyMonkey was founded in 1999 by Ryan and Chris Finley in Portland. For years it was the undisputed leader and remains one of the most widely used platforms, especially for simple commercial surveys.

Typeform was founded in Barcelona in 2012 by David Okuniev and Robert Muñoz. Its differentiator was the experience design (one question at a time, smooth transitions, a more conversational format) that significantly improved completion rates in online surveys. It's one of the Spanish SaaS unicorns. It has been acquired in private equity deals in recent years.

Qualtrics, founded in 2002 in Provo, Utah, specialized in rigorous research for enterprise and academia. It went public in 2018, was acquired by SAP in 2018 for 8 billion dollars, and was taken private again (in 2023) after an ownership transition.

Google Forms is the most widely used free option for internal, simple, and non-critical surveys. Zero cost, integration with Google Workspace, enough for basic needs.

Tally, Jotform, Formstack, and Microsoft Forms are alternatives with specific positionings.

For qualitative research specifically: Dovetail, Hotjar, and Lookback offer analysis and interview capabilities that complement surveys with deeper research.

The choice of tool matters less than the methodology. A poorly designed survey in Qualtrics produces garbage data just as much as one poorly designed in Google Forms.

How to design a questionnaire that produces useful data

The practices that distinguish rigorous questionnaires from those that generate pretty noise:

Define the decision you're going to make before writing questions. If you won't make any different decision after the results, don't run the survey. Any question whose answer changes nothing is noise.

Limit the length aggressively. Completion rates fall exponentially with length. Five-minute surveys get finished; fifteen-minute ones don't. Every extra question must justify its cost in abandonment. If you're unsure whether to include a question, don't include it.

Start with easy questions, leave the hard ones for the middle. Demographic questions at the beginning or end (not in the middle, to avoid fatigue). Controversial or intimate questions in the middle, when there's engagement but before fatigue sets in.

Avoid double-barreled questions. "Do you find our product fast and easy to use?" mixes two attributes. Separate them.

Mind the language without biasing it. "Do you agree that our innovative product solves your problem?" contains several assumptions (that it's innovative, that you have the problem). Ask it neutrally.

Offer a "not applicable" or "don't know" option. Forcing a response when there's no opinion generates false data.

Randomize the order of options when applicable. Order affects selection (primacy/recency effect). Randomizing avoids systematic bias.

Pre-test with 5-10 people before launching. They'll find problems you don't see.

Be transparent about data use. Privacy, anonymization, purpose. Trust in how the data is used influences the honesty of responses.

Combine open-ended with closed questions. Closed ones give aggregable numbers; open ones give nuance you didn't anticipate. A few open-ended questions (1-3 per questionnaire) are enough.

Common mistakes in questionnaires

Running a survey with no clear objective. More common than it seems. People launch surveys because it was time, not because there was a decision to make.

Confusing the sample with the audience. Those who respond represent those who respond, not your general audience.

Asking about future intention. "Would you pay for X?" is the most deceptive question there is. People dramatically overestimate what they would do hypothetically. Ask about past behavior, not future.

Assuming representativeness. A survey of current customers shows current customers. If you want to know about prospects or non-customers, the methods are different.

Over-interpreting small differences. If NPS went from 32 to 35, it doesn't mean something big changed. Statistical variability is usually larger than small differences.

Ignoring non-response rates. If you publish a result without reporting the response rate, you're hiding information that's critical to interpretation.

Misusing NPS or standard metrics. NPS works as an internal trend. Comparing your NPS with that of another industry, another company, or another geography is usually misleading because the baselines are different.

Not acting on the results. A questionnaire that produces findings that aren't applied is a decorative questionnaire. Close the loop.

Questionnaires and creative operations

For a brand that wants to feed creative production with customer research, questionnaires are critical input for almost every decision: what content to produce, what message works, what's failing in the funnel. But they're only useful when the results travel to the creative team and translate into decisions. A report that gets filed away affects nothing.

That translation is the discipline of creative operations: the questionnaire findings feed the editorial calendar (which topics to address), the empathy map (which motivations to document), the creative KPIs (what to measure as success after the change).

At Polimake the research inputs —including questionnaire results— ideally live in accessible Media, Studio coordinates the derived production, and Studio executes the changes. Without that circuit, questionnaires produce pretty presentations that no one applies.

---

If you lead research, marketing, product, or strategy and you've gotten here looking for an answer about questionnaires, the most useful thing you can take away from this article is probably the combination of three ideas: sample size matters less than representativeness (the lesson of Gallup vs. Literary Digest, still valid ninety years later), response rates have fallen so far that many commercial surveys aren't representative of anything (5-10% typical today versus 80% in the 1970s), and a questionnaire with no decision attached is wasted work. Designing a serious survey is methodological work, not just writing questions.

To round it out, commercial research techniques covers the broader context in which the questionnaire is one tool among several, empathy map covers how findings translate into a customer profile, and market research covers the general discipline.

Quick references

• Commercial research techniques — the broader context.
• Market research — the general discipline.
• Empathy map — how to translate findings into a customer profile.
• Active listening on social media — the complementary qualitative method.
• Competitive analysis — different but adjacent research.