Retour

How to Code Open-Ended Survey Responses: The Market Research Workflow

Matthieu Saussaye

Start here: SmartInterview

Coding open-ended responses means turning free-text answers into numbered codes you can count, crosstab and track. The manual way takes a codeframe, trained coders and a quality check. SmartInterview does it AI-first: it collects the answers, probes vague ones with an AI follow-up question, then codes them into themes and sub-codes with sentiment, while you review and edit.

Try SmartInterview free: your first 50 respondents are free, no credit card required.

The Short Answer

To code open-ended survey responses, you build a codeframe from a sample of answers, group codes into nets, write coding rules, code every answer, double-code a sample to check consistency, then deliver the codes as variables in your tables. The steps:

  1. Define the output: which tables and variables the coding must feed

  2. Build the codeframe from a sample of real answers, with numbered codes

  3. Group codes into nets so you can report at two levels

  4. Write coding rules: multi-coding, "other", vague answers, minimum mentions

  5. Code the full dataset, manually, AI-assisted or AI-first

  6. Double-code a sample and resolve disagreements

  7. Deliver codes as multi-response variables, with example quotes per code

This guide is the production workflow used on market research projects. If you are deciding between interviews or focus groups and a survey, start with qualitative vs quantitative research. If your problem is answers too short to code, fix the question first with our guide to open ended questions.

Manual, AI-Assisted and AI-First Coding Compared

There are three ways to run the same workflow. Dedicated coding tools such as Codeit and Ascribe Coder now support all three, from fully manual to largely automatic, according to their vendors. The difference is who does the first pass and where the answers come from.

Step

Manual (spreadsheet or coding tool)

AI-assisted (Codeit-style tool)

AI-first (SmartInterview)

Collecting answers

In a separate survey platform

In a separate survey platform, synced or imported

In the same tool, with AI follow-up probing and voice answers

Codeframe

A coder reads a sample and drafts it

AI extracts a draft codeframe, a coder refines it

AI builds themes and sub-codes, you edit them

Coding every answer

Coder by coder, answer by answer

AI autocodes, coders review; past waves train the model

Automatic as answers arrive

Quality check

Second coder on a sample

Second coder plus the tool's QC features

Review quotes per code, edit, recheck a sample

Delivery

Codes typed into the data file

Codes written back to the survey or exported

Dashboard, plus Excel or SPSS export

Best for

Small datasets, teaching

Agencies, trackers, coding teams

Teams that run their own studies

For a side-by-side of the coding tools themselves, see Codeit vs Ascribe vs Caplena vs Displayr.

Code your next open question automatically, free · Ask for a quote

Step by Step: Coding Open-Ended Responses

1. Start from the deliverable

Before reading a single answer, decide what the coding must produce. Usually that is one multi-response variable per open question, where each respondent can carry several codes, plus net variables. Agree the format with whoever builds the tables. A codeframe that cannot be read by the tables is rework.

2. Build the codeframe from a sample

Read a sample of answers, a few hundred on a large study or all of them on a small one, and list the distinct ideas. Then turn the list into codes with a number, a short label and a one-line definition. A good codeframe is:

  • Grounded: every code comes from what people actually wrote, not from the brief

  • Mutually exclusive: two codes never mean the same thing

  • Exhaustive: it includes "Other", "Don't know" and "Nothing / no comment"

  • Specific enough to act on: "Price too high vs competitors" beats "Price"

3. Group codes into nets

Nets are the summary level: a net counts respondents who mentioned at least one code inside it. A net is not the sum of its codes, because one respondent can mention two codes in the same net and is counted once. Typical nets separate positive and negative mentions of the same topic.

Code

Label

Net

100

NET: Price (negative)


101

Too expensive overall

Price (negative)

102

More expensive than competitors

Price (negative)

103

Price increase / hidden fees

Price (negative)

200

NET: Ease of use (positive)


201

Easy to set up

Ease of use (positive)

202

Intuitive interface

Ease of use (positive)

997

Other


998

Don't know


999

Nothing / no comment


Leave gaps in the numbering (101, 102, 103, then 200) so new codes can be added later without renumbering, which matters a lot on trackers.

4. Write the coding rules

Rules make two coders reach the same answer. Write them down in the codeframe file:

  • One answer can take several codes; code every distinct idea

  • When an answer is too vague for a specific code, use the net-level "general" code instead of guessing

  • Codes below a minimum share of mentions are merged into a broader code or into "Other"

  • "Other" should stay small; if it grows, review it and create new codes

  • Brand mentions are coded against a fixed brand list, with misspellings mapped to the right brand

5. Code the full dataset

Manually, a coder reads each answer and applies codes. In an AI-assisted tool, the AI does the first pass and the coder reviews. Codeit, for example, says its Extract feature builds a codeframe and autocodes against it, guided by your research objectives and question text, and that it can work with your required themes, nets or an existing codeframe. In an AI-first platform like SmartInterview, coding runs automatically on each open question as answers arrive, and you edit the themes and sub-codes rather than build them.

Use the "Open-ended question" template: an open question with AI follow-up probing, coded automatically into themes and sub-codes.

6. Double-code a sample

Have a second coder, or a second pass, code a random sample independently. Compare the two. Where they disagree, the cause is usually a vague code definition or a missing rule, so fix the codeframe and recode the affected answers. Report the agreement rate in your methods note. The same check applies to AI output: treat the AI as one coder and a human as the second.

7. Deliver codes, counts and quotes

Export one variable per code (or a multi-response set) plus the nets, so the tables can crosstab them by segment. Always keep two or three verbatim examples per code. A share like "18% mention price" is only convincing next to the words people used.

Multilingual Open-Ended Responses

Multi-country studies add one decision: code in the original language, or translate first and code in one language.

  • Translate, then code: one codeframe, one coding team, consistent codes across markets. The risk is losing nuance in translation, so keep the original next to the translation. Codeit, for example, says it translates 130+ languages on import and shows original and translated verbatims side by side.

  • Code in the original language: native coders catch nuance, but you need one master codeframe translated into each language and a check that codes are applied the same way everywhere.

  • Collect and code in one multilingual flow: SmartInterview runs the survey in each respondent's language and codes all answers into one codebook, so you compare markets on the same codes.

Whatever the route, one rule holds: one master codeframe, with the same code numbers in every language.

Coding Trackers and Repeated Waves

On a tracker, consistency matters more than perfection. Freeze the codeframe after wave one. Add new codes with new numbers, never reuse or renumber old ones, and note in the codeframe the wave in which each code was added. Codeit says its machine learning learns from past coded waves to autocode new ones in the same way; Ascribe's Coder offers a Lightning Mode that suggests codes for uncoded verbatims from already coded ones.

Use Cases

Three typical situations, as illustrations rather than client stories.

NPS: coding the why behind the score

Code promoters' and detractors' comments separately with the same codeframe, then compare nets. The gap between the two is your driver list. Probing the score with a follow-up question gives you reasons instead of "good service".

Use the "NPS deep-dive with driver ratings" template: the 0 to 10 likelihood score, three AI follow-ups on the why and the how, and four driver ratings.

Churn: coding cancellation reasons

Exit answers are short and emotional. Use nets for price, product gaps, service and change of need, and a code for "switched to a competitor" with the competitor named. A follow-up question on what would have kept the customer gives the most actionable codes.

Use the "Cancellation exit interview" template: a survey to send right after a customer asks to cancel, built to capture the real reason.

Coding a file you already have

Last year's open answers sit in a spreadsheet. In SmartInterview you can upload a file of open answers to code it, then compare its themes with a new study. For more on getting depth out of AI surveys, read how to get more insights from AI surveys.

Common Mistakes When Coding Open-Ended Responses

  • Building the codeframe from the brief. Codes nobody mentioned stay empty, and real themes end up in "Other".

  • Codes that are too broad. "Quality" or "Service" cannot be acted on. Split them.

  • Adding net counts. Nets count respondents, not mentions.

  • Skipping the second coder. Without a consistency check, nobody can tell whether a shift between waves is real or a coding drift.

  • Trusting AI output unread. AI coding is a strong first pass, but read the quotes behind each code before you report it.

  • Coding answers that were never probed. If most answers are two words long, the fix is at collection, not in the codeframe.

Choosing a tool for this? See SmartInterview, the Codeit alternative, Codeit pricing 2026 and the Caplena alternative page.

Collect and code open answers in one place, free for 50 respondents →

Sources (checked November 8, 2026): Codeit Extract; Codeit Refine; Codeit Apply; Codeit Auto Translation; Ascribe Coder.

Frequently Asked Questions

How do you code open-ended survey responses?

Build a codeframe from a sample of answers, group codes into nets, write coding rules, code every answer, double-code a sample to check consistency, and export the codes as variables for your tables, with example quotes per code.

What is a codeframe in market research?

A codeframe is the numbered list of codes used to classify open answers, each with a label and a definition, usually grouped into nets. It is the market research name for a codebook.

What is a net in survey coding?

A net is a summary code that counts respondents who mentioned at least one code in a group. Because one person can mention several codes in the same net, a net is not the sum of its codes.

How many open-ended answers should you read to build a codeframe?

Enough that new answers stop adding new codes. On a large study that is often a few hundred answers; on a small one, read them all. Keep "Other" small and revisit it.

Can AI code open-ended responses?

Yes. Tools like Codeit and Ascribe offer AI-assisted coding, and SmartInterview codes open answers automatically into themes and sub-codes. In every case, review the quotes behind each code and double-code a sample before reporting.

Should multilingual answers be translated before coding?

Either works if you keep one master codeframe with the same code numbers in every language. Translating first is simpler to manage; coding in the original language keeps more nuance.

Table des matières

Articles associés

Inscrivez-vous gratuitement

Inscrivez-vous gratuitement

Inscrivez-vous gratuitement