Back

How to Conduct a Survey: 9 Steps From Goal to Report

Matthieu Saussaye

To conduct a survey, start from the decision the results must inform, define exactly who you need to hear from, decide how many responses you need, write and pretest a short questionnaire, field it through the right channel, then clean, analyze and report against your original question. Skipping any of these steps is how surveys end up producing numbers nobody uses.


The nine steps, in order:

  1. Write the decision and the research question. One sentence each. If you cannot name the decision, do not launch yet.

  2. Define the target population. Who counts, who is excluded, and how you will screen for them.

  3. Choose the method and channel. Online, phone, in person, in app, voice or text.

  4. Decide the sample size and the subgroups you need to read.

  5. Draft the questionnaire. Map every question to an objective; cut the rest.

  6. Pretest it. Five to ten real people, watched or interviewed afterwards.

  7. Field the survey. Invitation, reminders, quotas, live quality monitoring.

  8. Clean and analyze. Remove speeders and duplicates, code open answers, cut by subgroup.

  9. Report and act. Lead with the answer to the research question, then tell respondents what changed.

The rest of this guide walks through each step with the choices you actually face, a planning table, the mistakes that cost the most, and a one-page checklist.

Survey methods compared: which one fits your study

Method is the choice that most constrains cost, speed and what the data can tell you. Pick it before you write questions, because a question that works on a laptop screen can fail on the phone or out loud.

Method

Best for

Speed

Depth of answers

Main risk

Online self-completion (email or link)

Customers, employees, panels; most commercial studies

Days

Low to medium; open answers tend to be short

Breakoff and low-effort answering on long forms

In-app or website intercept

Feedback tied to a moment of use

Hours to days

Low; keep it to a few questions

Only reaches active users, so it skews positive

Phone (CATI)

Older or offline audiences, B2B decision makers

Weeks

Medium to high; interviewers can probe

Cost per interview and interviewer effects

Face to face

Hard-to-reach groups, product handling, long interviews

Weeks to months

High

Highest cost, social desirability bias

AI-moderated voice or text interview

Open, exploratory questions at survey scale

Days

High; follow-up questions on every answer

Needs good probing rules so follow-ups stay on topic

Paper / mail

Populations with an address list but low web use

Weeks

Low to medium

Slow, and data entry adds errors

Most studies today run online, sometimes mixed with phone for groups that are hard to reach by email. If you need to understand why people do something, not only how many do it, plan for open questions and a way to probe them. Our comparison of data collection methods goes deeper on each channel.

Step 1: Start from the decision, not the questionnaire

The most common reason a survey disappoints is that it was written as a list of interesting questions rather than as an answer to a decision. Before anything else, write two sentences:

  • The decision: "We will decide whether to launch the premium tier in Germany in Q2."

  • The research question: "What share of current German customers would pay 30% more for priority support, and what would make them hesitate?"

Then list the three to five things you must learn to answer that research question. These become your objectives, and every question in the final survey should trace back to one of them. A question that maps to no objective is a cost to the respondent with no return for you.

Also write down what result would change your mind. If the team would launch whatever the data says, the survey is a formality and you can save everyone's time.

Step 2: Define who you need to hear from

"Our customers" is not a population definition. Specify it precisely enough that two people would agree on whether a given person qualifies: country, age range, customer status, purchase recency, role in the buying decision. Then decide how you will identify them in practice, which usually means two or three screening questions at the start.

Write down who you are excluding as well, for example employees, people who bought only once more than two years ago, or respondents who work in market research. Exclusions are what keep a sample clean. If you still need to work out where these people come from, our guide on how to find survey respondents compares customer lists, intercepts and panels.

Step 3: Choose the method and channel

Use the table above as a starting point, then check three practical constraints:

  • Can you reach the population on that channel? An email survey of people whose email you do not have is not a plan.

  • What device will they answer on? If most will answer on a phone, avoid wide grids and long scrolling pages.

  • Do you need explanations or only counts? Counts suit closed questions online. Explanations need open questions and some form of probing, by an interviewer or by an AI moderator.

Step 4: Decide how many responses you need

Sample size depends on the precision you need and, more importantly, on the smallest subgroup you plan to report on. If you want to compare three regions, each region needs enough responses on its own. A total of 400 split unevenly across three regions may leave one region with 60, which is too few to say much.

For a quick estimate of the margin of error at 95% confidence for a proportion near 50%, the standard formula is 1.96 × √(0.25 / n). That gives roughly ±10 points at n = 100, ±5 points at n = 400 and ±3 points at n = 1,000. Use our sample size calculator for your own numbers, and remember that this margin only strictly applies to random samples. Panel and convenience samples should be described by their method, not dressed in a margin of error.

Then work backwards from your expected response rate: if you need 400 completes and expect 10% of invitees to finish, you need about 4,000 invitations, plus extra for screen-outs.

Step 5: Write the questionnaire

Write in the order the respondent will experience it:

  1. A short introduction: who is asking, why, how long it takes (honestly), and how the data is used.

  2. Screeners: the two or three questions that confirm the person belongs in the study.

  3. Easy, general questions first, then the specific and harder ones.

  4. The core objectives, each with a mix of closed questions for counting and one open question for the why.

  5. Sensitive and demographic questions last, once trust is established.

  6. A thank-you screen that says what happens next.

Question-writing rules that prevent most problems:

  • One idea per question. "How satisfied are you with our price and delivery?" is two questions.

  • Neutral wording. "How would you rate..." not "How much did you enjoy...".

  • Balanced scales with a clear midpoint and labelled end points.

  • A "Don't know" or "Not applicable" option where people genuinely may not know.

  • Specific time frames. "In the last 30 days", not "recently".

Keep it short. SurveyMonkey's analysis of about 100,000 surveys found that respondents spend roughly 5 minutes on a 10-question survey, and that abandonment rose for surveys taking more than 7 to 8 minutes, with completion rates dropping 5% to 20% (SurveyMonkey, How long should a survey be?).

Step 6: Pretest before you launch

A pretest is the cheapest quality control in research. Send the survey to five to ten people who match the target population, not colleagues who already know what you meant. Ask them to think aloud as they answer, or call them afterwards and ask what they understood each question to mean.

You are looking for questions people interpret differently from you, answer options that do not fit their situation, skip logic that sends them to the wrong place, and the point where they started to get bored. Also time it: the honest length goes in the invitation. Then run a soft launch to the first 5% to 10% of your sample and check the data before releasing the rest.

Step 7: Field the survey

Fielding is where most of the response rate is won or lost. Three things matter most:

  • The invitation. Personal sender, a specific reason, the real length, and the link above the fold. Keep the invitation to five sentences or fewer.

  • Reminders. One or two, a few days apart, only to non-responders. In a Pew Research Center experiment, adding text messages to email reminders produced faster responses (48% vs 39% response by day 10) but no significant difference by the end of the field period (Pew Research Center, 2016). Reminders mostly buy speed; plan a field period long enough that you do not need to over-remind.

  • Live monitoring. Watch quota cells fill, check the dropout point by question each day, and read the first open answers. Fixing a broken question on day one saves the whole study.

Step 8: Clean and analyze the data

Before you calculate anything, remove responses that should not count:

  • Speeders: completes far faster than a plausible reading time, often below a third of the median duration.

  • Straightliners: identical answers down every grid, especially where some items are reversed.

  • Duplicates and bots: repeated IP or device fingerprints, gibberish in open answers.

  • Failed screeners: people who contradict their own qualifying answers later in the survey.

Then analyze in layers: top-line results for each objective, the same results cut by the subgroups you planned in step 4, and the open answers coded into themes so you can count them. Coding open answers is the step most teams skip because it is slow by hand; our guide to open-ended questions covers how to write and code them.

Step 9: Report, act and close the loop

Lead the report with the answer to the research question from step 1, in one sentence, before any chart. Then give the evidence for each objective, the subgroup differences that matter for the decision, and a short note on method: who was asked, how, when, how many responded, and what that means for how far the results can be generalized.

Finally, tell respondents what you learned and what you will do about it. People who see that a survey changed something are more willing to answer the next one; people who never hear back learn that answering is pointless.

Common mistakes when conducting a survey

Mistake

What it causes

Fix

No decision behind the survey

Interesting results nobody acts on

Write the decision and research question first

Population defined loosely

Mixed sample, results you cannot interpret

Precise inclusion and exclusion rules plus screeners

Too few responses per subgroup

Differences you cannot trust

Size the sample for the smallest group you report

Questions nobody pretested

Misread questions, broken skip logic

Five to ten cognitive pretests plus a soft launch

Length understated in the invite

Breakoff mid-survey and lost trust

Time it in pretest and state the real number

Open answers left unread

The why behind the numbers is lost

Code open answers into themes and count them

Most of these mistakes also introduce survey bias that no amount of analysis can remove afterwards.

A one-page survey checklist

  • Decision and research question written in one sentence each

  • Three to five objectives, each mapped to specific questions

  • Population, exclusions and screener questions defined

  • Method and channel chosen for the device people will use

  • Sample size set for the smallest subgroup you will report

  • Questionnaire under 10 minutes, demographics at the end

  • Pretested with 5 to 10 target respondents, then soft launched

  • Invitation with honest length, plus one or two reminders scheduled

  • Cleaning rules written before you see the results

  • Report leads with the answer, ends with method and next steps

Run the whole process with depth built in

The steps above stay the same whatever tool you use. What changes is how much you learn from each respondent. In most online surveys, the open question that explains the numbers gets a few words, and nobody codes them.

SmartInterview runs the survey as a conversation, by voice or text, and an AI follows up on vague answers in the respondent's own language. Quotas keep your subgroups on target while it fields, and open answers come back automatically coded into themes you can count next to the closed questions.

Start free

Frequently Asked Questions

What are the steps to conduct a survey?

The core steps are: define the decision and research question, define the target population, choose the method, set the sample size, write the questionnaire, pretest it, field it with reminders, clean and analyze the data, and report the results. Guides list between five and twelve steps, but they all cover this same sequence.

How long should a survey be?

Aim for under 10 minutes, and shorter for customer surveys. SurveyMonkey's data shows abandonment rising for surveys longer than 7 to 8 minutes, and respondents spending less time per question as surveys grow. Work, school and panel audiences tolerate more length than customers do.

How many people do I need to survey?

It depends on the precision you need and the smallest subgroup you will report. As a rough guide for a random sample, 100 responses give about ±10 points, 400 give about ±5 points and 1,000 give about ±3 points at 95% confidence. Size each subgroup you plan to compare, not only the total.

What is the best way to conduct a survey?

For most audiences, an online survey sent through a channel they already use, kept short, pretested, and followed by one or two reminders. If you need to understand reasons, add open questions and a way to follow up on them, through interviewers or an AI moderator.

How do you analyze survey results?

Clean the data first by removing speeders, straightliners, duplicates and failed screeners. Then report the top-line answer to each objective, cut it by the subgroups you planned, and code open answers into themes so they can be counted alongside the closed questions.

Should I pretest a survey?

Yes, always. Five to ten people from the target population reveal most wording problems, wrong answer options and broken skip logic. A soft launch to the first 5% to 10% of the sample then catches technical issues before the full budget is spent.

Table of contents

Related articles

Sign up for free

Sign up for free

Sign up for free