Must a close-ended question have only two answers?

No. It can offer multiple categories, an ordered scale or several selectable options.

Should every question include other or do not know?

Not automatically; include options that fit the measurement purpose and respondent experience, and test them.

Is a rating from one to five a precise numerical measurement?

It records an ordered response, but equal numerical spacing does not by itself prove equal differences in the underlying attitude.

AI Glossary / Statistics and probability / Beginner

Close-ended question

Definition

A close-ended question is a question that asks respondents to choose from a defined set of answers rather than compose an unrestricted response. It is also commonly called a closed-ended question. Examples include yes/no items, rating scales and multiple-choice questions. In surveys about AI, the available options shape what can be measured, so designing those options is part of designing the evidence itself.

What it means

The researcher decides the response framework in advance. Respondents might select one category, select several applicable categories or give a rating on a specified scale. This structure makes answers easier to compare and summarize, but limits what respondents can express. An answer can reflect the offered categories, their wording and their order as well as the person's underlying view. A missing option can force an inaccurate answer, produce nonresponse or hide an important experience that an open-ended question might have revealed.

How it works

Start with a clear measurement objective and a question that asks about one thing at a time. Define the relevant time period and terms. For single-choice categories, avoid overlap and cover plausible answers, including appropriate uncertainty or non-applicability options. Explain whether respondents should choose one answer or all that apply. Use balanced, clearly labeled scales rather than assuming that everyone interprets bare numbers identically. Test the wording with intended respondents. Consider order effects and randomization where suitable, while keeping inherently ordered scales understandable. During analysis, distinguish valid answers, skipped items and responses such as do not know.

Three practical examples

AI adoption survey: ask whether a person used an approved assistant during the past week, with explicit yes, no and uncertain options.
Service evaluation: ask users to rate the clarity of a generated explanation on a fully labeled response scale.
Feature research: present specified tasks and ask which the respondent performed, making the multiple-selection instruction explicit.

Think of it this way

Think of a restaurant order form with checkboxes. It makes orders easy to count, but only for dishes and modifications the form allows people to describe. If a needed option is absent, a tidy spreadsheet can still misrepresent what customers wanted. Clear categories matter as much as convenient counting.

Worked example

A hypothetical employee survey asks, "Did you use the approved AI assistant for a work task in the past seven days?" Among 120 submitted questionnaires, 72 answer yes, 28 answer no, 10 choose unsure and 10 leave the item blank. Yes responses are 72/120 = 60% of all submitted questionnaires. Among the 100 respondents who answered either yes or no, the yes percentage is 72/100 = 72%. Neither denominator is interchangeable with the other. A transparent report gives the counts and explains which population each percentage describes. Treating unsure and blank as no would create 48 apparent no responses, although only 28 people actually selected no. The calculation cannot establish the adoption rate for employees who never submitted a questionnaire.

Illustrative case study

Fictional case: Brookside Services surveys staff about an internal assistant. Its first draft asks whether the tool is "fast and useful," with only agree and disagree responses. Test participants explain that it is fast for some tasks but unreliable for others. The team separates speed from usefulness, specifies the period being recalled and adds a non-use option where appropriate. It also includes an optional text box for issues outside the response list. The resulting survey remains easy to summarize while giving respondents a more accurate way to describe their experience. This invented case demonstrates questionnaire improvement, not measured evidence about a real workplace.

Common misconceptions and mistakes

Closed-ended does not mean unbiased or automatically objective. Agree/disagree wording can encourage acquiescence, and an unbalanced scale can push responses in one direction. Select-all-that-apply and separate yes/no items can produce different results even when they mention the same behaviors. Numeric response labels do not automatically create equal distances between categories. Do not treat a non-applicable answer as the lowest satisfaction score. Double-barreled questions, unfamiliar technical language and unspecified recall periods weaken interpretation. Large samples cannot repair a question that measures something different from the intended concept.

Business use

Businesses use close-ended questions to compare experiences, monitor adoption and identify common needs efficiently. For useful AI research, make the population, recall period and meaning of use explicit. Preserve the questionnaire wording and response options with the results so later teams can understand changes in measurement. If options change between survey waves, an apparent trend may reflect the instrument rather than behavior. Combine structured items with carefully chosen open-ended opportunities when unanticipated concerns would materially affect a product decision.

Check your understanding

Why does the worked example report both 60% and 72% without claiming a contradiction?

Answer: The first uses all 120 submitted questionnaires; the second uses only the 100 yes/no answers. Different denominators answer different questions. The uncertainty and missing responses must remain visible so readers can interpret both correctly.

Related terms

Open-ended question
Likert scale
Response Bias
Acquiescence bias
response option
questionnaire design
item nonresponse
recall period
survey denominator
cognitive interviewing

Learn next

compare single-choice, multiple-selection and rating formats. Practice rewriting double-barreled or leading questions, then test whether respondents interpret the revised item as intended. Study item nonresponse and denominator selection before comparing percentages across teams or survey waves. For changing AI terminology, verify that respondents understand the product or behavior named in the question.

Sources

NIST/SEMATECH, e-Handbook glossary, close-ended question: https://www.itl.nist.gov/div898/handbook/glossary.htm Pew Research Center, Writing Survey Questions: https://www.pewresearch.org/writing-survey-questions/ Pew Research Center, When Online Survey Respondents Only Select Some That Apply: https://www.pewresearch.org/methods/2019/05/09/when-online-survey-respondents-only-select-some-that-apply/

Previous: Clitic Next: Closed class