How to Code Open-Ended Survey Responses
Coding open-ended survey responses means assigning a defined category to what a respondent says. The result connects individual comments to a table you can analyse. A useful coding project preserves the original words, applies explicit rules and explains which respondents form the base of each percentage. You can do this in a spreadsheet before introducing any specialist software.
Start with the decision and the question
Write down what the analysis needs to inform. “Which product problems should we investigate?” calls for a different codebook from “Why did customers cancel?” Keep the exact survey question beside the answers. The response “nothing” can mean no improvement is needed, no benefit was received or no answer was offered. The question determines which interpretation is defensible.
Choose a coding unit. For a short survey comment, use the whole answer as the unit and permit several theme codes. You can identify meaning within individual phrases without splitting the respondent into several records. Keep separate questions in separate coding projects or clearly named variable groups; identical words may mean different things across questions.
Prepare a spreadsheet you can trace
Create a read-only raw sheet with respondent ID, question ID and original text. Work in a copy. Retain IDs as text when leading zeros matter, and check that each row still matches its source after sorting. Remove contact details from the working text where they are unnecessary for analysis. Keep any permitted link back to identifiable records separately.
Add one column per theme, using 1 for mentioned and 0 for not mentioned after review. Add a response-status column and a notes column. A blank code cell should mean unfinished work, rather than silently turning into zero. Store “not asked” separately if survey routing explains the absence of an answer.
Build and pilot the categories
Read a spread of answers: short and long comments, different customer groups, and different collection periods. Draft categories that correspond to the analysis question. Give each a stable ID, a label, a definition and boundaries. “Product” is usually too broad to help a product manager; a separate code for every wording variation is usually too narrow to report.
Use a small pilot batch to expose overlaps. Code it, record difficult decisions and revise the definitions before coding the remaining rows. If another analyst is available, have them apply the draft independently and discuss the reasons for disagreements. If working alone, revisit the pilot later without looking at the first pass. Neither process justifies claiming that every remaining answer is correct.
Worked example: feedback about a subscription app
These ten invented answers respond to “What should we improve about the app?” They are examples, not research findings. The companion CSV codebook contains the full upload fields.
| Code | Apply when the answer mentions | Boundary |
|---|---|---|
| ease | Learning, setup or navigation | Visual appearance alone belongs to design |
| reliability | Crashes, errors or loading speed | A missing capability belongs to features |
| price | Price, billing or value | Effort alone is not a monetary cost |
| features | A capability, integration or export | Difficulty finding an existing tool belongs to ease |
| support | Contact with help or service staff | General product faults do not imply support contact |
| design | Appearance, typography or layout | Navigation difficulty alone belongs to ease |
| ID | Answer | Assigned codes or status |
|---|---|---|
| A01 | “I get lost finding my saved reports.” | ease |
| A02 | “It crashes when I open a large report.” | reliability |
| A03 | “The monthly charge is too high for occasional use.” | price |
| A04 | “Please let us export charts to PowerPoint.” | features |
| A05 | “Support sent the same canned reply twice.” | support |
| A06 | “The grey text is difficult to read.” | design |
| A07 | “Finding invoices is confusing, and the charge keeps rising.” | ease; price |
| A08 | “It never crashes; I only want an offline mode.” | reliability; features |
| A09 | “Your company should sponsor our local sports club.” | other |
| A10 | “No answer.” | non_response |
A08 receives reliability because this code records mentions regardless of direction. It is a positive reliability statement inside an improvement question. If you only want reported problems, revise the code definition and recode all relevant answers. Do not change the interpretation for one inconvenient case.
Handle several themes without inflating respondents
A07 has two themes but remains one respondent. Within a theme, count the person once even if they repeat the complaint. Do not automatically add a broad parent category alongside a narrower child category unless the reporting design calls for both. If you calculate an overall service net later, derive it as the union of the relevant codes.
A09 is substantive but outside the current frame. A10 contains no usable response. Keep these statuses separate. Review the “other” bucket during the pilot: repeated substantive topics may deserve a new code, whereas unusual one-off suggestions may reasonably remain there.
Review both difficult and ordinary answers
Flag ambiguous wording, sarcasm, unclear pronouns and statements that need context. Write a short rationale when a decision sets a precedent. Also inspect a random selection of ordinary-looking assignments and non-assignments. Reviewing only flagged answers can miss systematic errors that nobody marked as difficult.
When a definition changes, identify which earlier answers could be affected and apply the new rule consistently. Save the codebook version alongside the completed dataset. Keep an unresolved status until you have decided; “other” should not become a convenient parking place for unfinished rows.
Turn codes into a transparent report
For each theme, report the number of respondents mentioning it. State whether the denominator is everyone asked, everyone who answered or everyone with a substantive comment. In this invented example, price appears in two answers; using ten submitted rows as the base differs from using nine substantive comments. Neither denominator should be hidden.
Explain that multiple themes can make percentages sum above 100 percent. Mention frequency does not establish importance, population prevalence or causation. A voluntary comment can reveal an issue worth investigating without showing how widespread it is. Use selected anonymised quotations to illustrate a defined category, and retain contrary examples when they matter to interpretation.
Before you hand over the file
- Confirm every source row is present and every ID is preserved.
- Resolve blank coding cells and distinguish missing text from other topics.
- Check multi-theme answers against the written rules.
- Reconcile code counts with the report and state its denominator.
- Save the raw data, final codebook, decisions and reviewed export together.
FAQ
Should I begin with categories or read the answers first?
Start with categories needed for the research question, then read actual answers to test their coverage. A hybrid approach avoids both an irrelevant predefined list and a frame driven entirely by memorable comments.
Can one answer receive contradictory codes?
It can mention contradictory experiences, such as helpful staff and an unresolved problem. Theme codes can coexist. If you also code direction, record which part is positive or negative rather than forcing the entire comment into one sentiment.
Should every answer receive a theme?
No. Preserve a distinction between a substantive answer outside the frame and a genuine non-response. Forcing either into a theme changes the meaning of the data.
Can I change a code halfway through?
Yes, with a recorded reason, a new version and a review of affected earlier answers. Otherwise the final table combines different definitions under the same label.
Survey Coding by Decision Models (Beta) supports CSV/XLSX uploads, your own codebook or a template, probabilities per code and review of uncertain answers. It exports CSV/XLSX and CSV with SPSS syntax, and is free for up to 500 answers per project. Try Survey Coding.