Survey Results Analysis
When To Use
The export is in and someone is about to average a Likert scale into a headline. Run it before any number from the survey reaches a slide. It answers: who answered, what the responses show with their bases, and what this data cannot say?
When Not To Use
If the questionnaire is not sent yet, use UX Survey Design first; good analysis cannot rescue a leading question. For interview transcripts, use Thematic Analysis. Survey open text is coded here.
Inputs
- The response export (spreadsheet or pasted table), the questionnaire wording, how people were invited, how many were asked, fielding dates
- Anonymise first: replace names with P1, P2 or response ids, remove contact details, employers and anything that identifies a person.
If you have none of this, I start from the export alone, say what I cannot know about who was asked, and mark the output as a first draft.
Approach
Survey analysis with base sizes and margin-of-error caution, from Pew Research Center's methods pages, Writing Survey Questions and Understanding the margin of error (2016): subgroups carry larger margins of error, and a change between surveys needs an even larger difference to mean anything. Most product surveys are convenience samples with no true margin of error, so a gap is a hint, not a finding. Distributions come before averages. The failure it prevents: a mean of the middle hiding one delighted group and one furious one, reported as "users are neutral".
Workflow
- Ask up to three questions: what decision does the survey serve, what exclusion rules do you set before we look at results, and what is the minimum base for a subgroup?
- State who answered: invited, answered, completed, fielding dates, channel. Set who answered against who was asked, and name who is likely missing.
- Write the cleaning log: each exclusion rule (speeders, straight-lining, duplicates, failed checks) with the count removed. Rules come from step 1, never invented after seeing the results.
- Frequencies per question with the base on every figure (n = [ ]). Rating scales as the full distribution, not an average; flag splits, floors and ceilings.
- Cross-tabs only where every subgroup meets the minimum base; below it, suppress the cell and say why. Call differences "hint" unless the sample design supports more.
- Code open text: codes with one-line definitions, counts as "[n] of [N] responses", verbatim quotes with response ids.
- Write what this data cannot say: why people answered as they did, what they will actually do, anything about non-responders. Turn each into a follow-up question for interviews.
Output Format
# Survey Results Analysis
**Survey:** [name] | **Invited:** [n] | **Answered:** [n] | **Completed:** [n] | **Fielded:** [dates] | **Channel:** [how reached] | **Likely missing:** [who was asked but did not answer]
## Cleaning log
| Rule | Set before results | Removed |
|---|---|---|
| [speeders / straight-lining / duplicates / failed check] | [yes] | [n] |
## Frequencies
| Question | Answer distribution | Base (n) | Note |
|---|---|---|---|
| [question] | [each option with its count] | [n] | [split, floor, ceiling] |
## Cross-tabs
| Question | Subgroup | Base (n) | Distribution | Reading |
|---|---|---|---|---|
| [question] | [subgroup] | [n, or "below minimum, suppressed"] | [counts] | [hint / finding] |
## Open text codes
| Code | Definition | Count | Quotes (response id) |
|---|---|---|---|
| [code] | [one line] | [n] of [N] responses | "[verbatim]" (R[x]) |
## What this data cannot say
- [limit] -> follow-up question: [for interviews]
## Decision
[Research lead] decides which results go into the readout and which follow-up interviews run, by [date].
Done When
- Every figure carries its base (n = [ ]) from real responses
- Every exclusion rule is logged with a count and was set before results
- Subgroups below the minimum base are suppressed, and "what this data cannot say" is filled
Quality Bar
- No averages of rating scales as a headline; the distribution is shown
- No margin of error claimed for a convenience sample; differences there are hints
- No per-respondent profiles; subgroups below the minimum the user sets stay suppressed
- Open-text quotes are verbatim with response ids; no paraphrase in quotation marks
- Every figure carries its base from real responses; Claude never fills low response with estimates
Next
Run uxr-insight-statements (Research Insight Statements) to combine survey results with qualitative evidence.
About the makers
This pack is made by 90n (90n.ai): AI-native ways of working and tools for Product, Marketing and HR teams, built by ex-McKinsey founders. We map where time and money leak, redesign the work around AI and stay until your people run it themselves, and we run our own company on Claude. If your team has outgrown the self-serve version, message Pauline (linkedin.com/in/paulinebertry).
1---2name: uxr-survey-analysis3description: Analyses a survey export into who answered, a cleaning log, frequencies with base sizes, cross-tabs only where the base allows, coded open text and what the data cannot say. Use for "run uxr-survey-analysis", "analyse these survey results", "here is the survey export", "what do the numbers tell us", "Likert results", "cross-tab this survey", "code the open text answers", "is this difference real", part of the UX Research with Claude Pack by 90n.4metadata:5 source: https://skillmd.com/skills/polar-bear-org/uxr-survey-analysis6---78# Survey Results Analysis910## When To Use11The export is in and someone is about to average a Likert scale into a headline. Run it before any number from the survey reaches a slide. It answers: who answered, what the responses show with their bases, and what this data cannot say?1213## When Not To Use14If the questionnaire is not sent yet, use UX Survey Design first; good analysis cannot rescue a leading question. For interview transcripts, use Thematic Analysis. Survey open text is coded here.1516## Inputs17- The response export (spreadsheet or pasted table), the questionnaire wording, how people were invited, how many were asked, fielding dates18- Anonymise first: replace names with P1, P2 or response ids, remove contact details, employers and anything that identifies a person.19If you have none of this, I start from the export alone, say what I cannot know about who was asked, and mark the output as a first draft.2021## Approach22Survey analysis with base sizes and margin-of-error caution, from Pew Research Center's methods pages, Writing Survey Questions and Understanding the margin of error (2016): subgroups carry larger margins of error, and a change between surveys needs an even larger difference to mean anything. Most product surveys are convenience samples with no true margin of error, so a gap is a hint, not a finding. Distributions come before averages. The failure it prevents: a mean of the middle hiding one delighted group and one furious one, reported as "users are neutral".2324## Workflow251. Ask up to three questions: what decision does the survey serve, what exclusion rules do you set before we look at results, and what is the minimum base for a subgroup?262. State who answered: invited, answered, completed, fielding dates, channel. Set who answered against who was asked, and name who is likely missing.273. Write the cleaning log: each exclusion rule (speeders, straight-lining, duplicates, failed checks) with the count removed. Rules come from step 1, never invented after seeing the results.284. Frequencies per question with the base on every figure (n = [ ]). Rating scales as the full distribution, not an average; flag splits, floors and ceilings.295. Cross-tabs only where every subgroup meets the minimum base; below it, suppress the cell and say why. Call differences "hint" unless the sample design supports more.306. Code open text: codes with one-line definitions, counts as "[n] of [N] responses", verbatim quotes with response ids.317. Write what this data cannot say: why people answered as they did, what they will actually do, anything about non-responders. Turn each into a follow-up question for interviews.3233## Output Format34```markdown35# Survey Results Analysis36**Survey:** [name] | **Invited:** [n] | **Answered:** [n] | **Completed:** [n] | **Fielded:** [dates] | **Channel:** [how reached] | **Likely missing:** [who was asked but did not answer]37## Cleaning log38| Rule | Set before results | Removed |39|---|---|---|40| [speeders / straight-lining / duplicates / failed check] | [yes] | [n] |41## Frequencies42| Question | Answer distribution | Base (n) | Note |43|---|---|---|---|44| [question] | [each option with its count] | [n] | [split, floor, ceiling] |45## Cross-tabs46| Question | Subgroup | Base (n) | Distribution | Reading |47|---|---|---|---|---|48| [question] | [subgroup] | [n, or "below minimum, suppressed"] | [counts] | [hint / finding] |49## Open text codes50| Code | Definition | Count | Quotes (response id) |51|---|---|---|---|52| [code] | [one line] | [n] of [N] responses | "[verbatim]" (R[x]) |53## What this data cannot say54- [limit] -> follow-up question: [for interviews]55## Decision56[Research lead] decides which results go into the readout and which follow-up interviews run, by [date].57```5859## Done When60- Every figure carries its base (n = [ ]) from real responses61- Every exclusion rule is logged with a count and was set before results62- Subgroups below the minimum base are suppressed, and "what this data cannot say" is filled6364## Quality Bar65- No averages of rating scales as a headline; the distribution is shown66- No margin of error claimed for a convenience sample; differences there are hints67- No per-respondent profiles; subgroups below the minimum the user sets stay suppressed68- Open-text quotes are verbatim with response ids; no paraphrase in quotation marks69- Every figure carries its base from real responses; Claude never fills low response with estimates7071## Next72Run uxr-insight-statements (Research Insight Statements) to combine survey results with qualitative evidence.7374## About the makers7576This pack is made by 90n ([90n.ai](https://90n.ai)): AI-native ways of working and tools for Product, Marketing and HR teams, built by ex-McKinsey founders. We map where time and money leak, redesign the work around AI and stay until your people run it themselves, and we run our own company on Claude. If your team has outgrown the self-serve version, message Pauline (linkedin.com/in/paulinebertry).