How to analyze survey data: from raw export to findings
How to analyze survey data well: clean the export, tabulate every question with its base, compare segments, code the open answers, report what matters.
Free app. Runs on your Mac with the Claude or Codex plan you already have.
What it means to analyze survey data
To analyze survey data is to turn a raw export of responses into a few findings a team can act on: what people answered, how groups differ, and why. Done well, every figure carries its base, the number of people it rests on, and the report opens with the result that changes a decision. Done badly, it is a deck of pie charts, with a headline built on a segment of nine people.
Survey tools chart each question and filter by a group. What they leave to you is judgment: which differences are real, what the open answers say and what to do about it. The method below is how an experienced research analyst works through a survey export by hand. Each step says what to do and why, then how 10folds carries it out.
What it does for the business
- Findings you can defendWhen every figure has its base and every difference passes a stated rule, the report holds up when someone asks how many people that was.
- See who differs, not just the averageSplitting results by plan, role or region shows where a problem is concentrated, which an overall score hides.
- Know why, not only whatCoding the open-ended responses links scores to the reasons behind them, so the team knows what to change.
- Avoid acting on noiseFlagging small groups stops a dramatic result from a handful of respondents turning into a project.
How to analyze survey data, step by step
Start from the purpose, then clean the data
Duplicates, speeders and half-finished responses distort every table that follows.
What to do
- Write down what the survey was for and the segments you need to compare.
- Remove empty and duplicate responses, and decide how to treat partial ones.
- Check that each question is read as the right type: rating, single choice, multi-select or open text.
- Recode skip codes such as not shown, so they are not read as answers.
- Note who was invited, when and how, to judge the response rate.
Why
Data cleaning comes first because every error carries into every table. A rating read as text, or a not shown code counted as an answer, changes the results quietly. Knowing who was invited tells you whether respondents resemble the people you care about. Writing the purpose down first also keeps the analysis focused on the questions that feed a decision, instead of every possible cut of the data.
Tabulate every question with its base
Sixty percent of ten people is six people.
What to do
- For each question, give the count and percentage for each answer, with the base.
- Treat a blank as unanswered, never as zero.
- Report multi-select questions as the share of people who answered, which can add up to more than 100 percent.
- For rating questions, give mean, median and spread.
- For a 0 to 10 recommend question, compute the Net Promoter Score: promoters minus detractors.
Why
The base size tells the reader how much weight a figure can bear. Counting blanks as zero lowers every average. The Net Promoter Score is widely used, but it hides the spread, so show promoters, passives and detractors beside it. Mean and median together show whether a few extreme ratings are pulling the average.
Compare segments with cross-tabulation
A difference on a small base is a signal to check, never a finding.
What to do
- Cross-tabulate each key question by the segments that matter: plan, role, company size, region.
- Set a minimum base below which a group is only a signal.
- Report a difference only when both groups meet the minimum and the gap is large.
- Use a significance test when the decision is costly, and judge whether the gap matters in practice.
- Never claim one thing caused another from a cross-tab.
Why
Cross-tabulation, a table of one question split by a group, is where most survey insights are found and most mistakes are made. Small groups produce large swings by chance. Statistical significance tells you whether a gap is likely real, but a real gap can still be too small to matter. Setting the minimum base before you look at the results stops you from lowering it to keep an interesting finding.
Code the open-ended responses
The reason behind a score is usually in the comment box.
What to do
- Read every open answer, not a sample.
- Write a short codebook: theme, definition, an example.
- Code each answer; one answer can carry several themes, and an other code is allowed.
- Count each theme and split it by your main segment.
- Pick two verbatim quotes per main theme.
Why
Open-ended responses explain the numbers: why detractors are unhappy, what a low rating means. Coding turns them into counts that sit beside the closed questions. Quotes make the theme real for readers who will not open the data. Reading every answer, rather than a sample, catches the rare but serious comment, such as a safety issue or a threat to leave.
Write the findings and check them
The first sentence of a survey report should be a finding.
What to do
- Open with the single most important finding, with its figure and base.
- Give three to five findings, each with figures, bases, open-text evidence, an action and your confidence.
- Add compact tables per question for readers who want detail.
- State the limits: responses removed, small segments, questions with many blanks.
- Have someone else check every figure against the tables.
Why
Readers decide from the first lines, so lead with what matters. Stating limits keeps the report honest about who answered. An independent check catches the most common error in survey reports: a figure quoted with the wrong base. Giving a confidence level with each finding tells readers which results to act on now and which to test further.
Common mistakes
- Skipping data cleaningDuplicates and skip codes change every result that follows.
- Hiding the baseA percentage without its base cannot be judged.
- Headlining a small segmentNine people swinging one way is not a trend; flag it and check it.
- Adding multi-select percentages to 100Each option is a share of people who answered.
- Reading cause into a cross-tabTwo things moving together does not mean one drives the other.
What a good result looks like
- A first sentence that states the main finding with its figure and base.
- Every percentage shown with its base.
- Segment differences reported only above your minimum base.
- Open-text themes with counts and verbatim quotes.
- A clear note of what was removed and who did not respond.
And then what? Using the result
- Share the main finding and its action with the team that owns it.
- Follow up small-segment signals with interviews or a larger sample.
- Chart the results from the tables for your own presentation.
- Rank the requests found in the open answers.
- Repeat the survey with the same questions to track change.
When to run it
- After a customer satisfaction or NPS surveyThe report shows the score, how it splits by segment and the reasons in the open answers.
- After a product or pricing surveyCross-tabs show which customers want what, with small groups flagged before anyone acts on them.
- When survey results sit unread in a toolA written report with findings first turns the export into decisions.
- Before presenting results to leadershipChecking every figure against its base before the meeting prevents the question that undoes a slide: how many people was that?
Questions people ask
How do you analyze survey data?
Clean the export, tabulate every question with its base, cross-tabulate the key questions by segment, code the open-ended answers, then report the few findings that change a decision, with their bases and limits.
What is cross-tabulation in a survey?
Cross-tabulation is a table that shows the answers to one question split by another variable, such as plan or region. It reveals how groups differ.
How do you analyze open-ended survey responses?
Read every answer, build a short codebook of themes with definitions, code each answer, count the themes and pick verbatim quotes for each main one.
How small a sample is too small to report?
It depends on the decision, but set a minimum base before you start and treat any group below it as a signal to check, not a finding. 10folds uses the smallest group you say you will trust.
Does 10folds weight the results?
No weighting is applied unless you supply it. Results describe the people who answered, not all your customers.
Does it contact respondents?
No. It works on your export only, and nothing is sent to respondents. Respondent IDs appear beside quotes so you can trace each one back.
More jobs in Product
Analyse survey responses, on your own Mac.Free app. Runs on your Mac with the Claude or Codex plan you already have. Your files stay on your computer.
Download for Mac