How to Find Common Themes in Survey Responses With AI

To find common themes in open-ended survey responses with AI, give the assistant clean, numbered responses, ask it to group related comments, then verify every theme and quote against the originals. A spreadsheet is enough to keep the findings traceable.

The responses to prepare before asking AI

A useful theme analysis starts with the original responses, not an AI summary of them. Put each answer in its own spreadsheet row and give each row a stable response ID. Keep the wording as submitted in one column. Use separate columns for the survey question and any information you are allowed to use, such as the service the respondent used. Remove names, contact details and other identifying information before sharing responses with an AI tool, and follow your workplace rules for handling survey data.

If you want to summarize Google Forms text responses with AI, open the form’s responses in a spreadsheet or export them, then select the open-ended question you want to examine. Keep answers to different questions separate at first. A comment about booking and a comment about delivery may use similar words but answer different questions.

Read through the responses yourself before sending them anywhere. Note blank entries, copied text and answers that do not address the question. Keep those rows in the working file so you can account for them, but mark them as unsuitable for theme assignment if appropriate. For a manageable set of responses, a spreadsheet and a general-purpose AI assistant are enough. You do not need a survey platform or an automated workflow to begin.

A prompt that makes themes traceable

To analyze open-ended survey answers with ChatGPT or another AI assistant, ask for themes tied to response IDs rather than a free-form summary. Paste the cleaned responses with their IDs and the survey question. If your workplace permits file uploads, a spreadsheet file can serve the same purpose. Check the tool’s data controls and your organisation’s policy before entering private feedback.

A practical prompt is: "Group these responses by the main issues they describe. Give each theme a plain-language name and a short definition. List the response IDs that support it. Allow a response to appear under more than one theme. Put unclear or unrelated responses in a separate group. Do not invent responses or rewrite text as a direct quote. Suggest possible representative quotes by ID only."

Read the output as a draft, not a finding. A theme named "communication" may sound plausible but cover unrelated problems. Ask the assistant to split broad themes into more specific ones, or merge themes that describe the same issue in different words. Keep the definitions and supporting IDs as you revise. Without them, a tidy list of themes gives you no reliable way to check what respondents actually said.

A worked example for grouping customer comments

Grouping customer survey comments by theme works best when the spreadsheet shows how each decision was made. Imagine a small service business reviewing open-ended feedback about making an appointment. The analyst keeps the original response in one column, its ID in another, and creates columns for proposed theme, final theme and review note. The actual customer wording stays in the original column throughout.

After reading the comments, the analyst might use "finding an available time" for comments about appointment options and "understanding the confirmation" for comments about what happens after booking. Those are proposed categories, not claims about any particular respondent. If one original response addresses both issues, both themes belong on that row. If the wording does not clearly support either, the analyst marks it for review instead of forcing a fit.

The AI assistant proposes theme names and supporting IDs. The analyst then opens each cited row and checks the definition against the original wording. If a cited response only mentions availability, it cannot support a finding about confirmation. When the labels are settled, the analyst filters by final theme to read all supporting comments together. This creates a usable result without relying on a made-up example quote or treating the assistant’s interpretation as evidence.

Representative quotes that remain faithful to the originals

A representative quote should be copied from an original response only after you have checked what the full response means. Ask the AI to suggest response IDs worth inspecting, not to produce polished quotations. Open each suggested row, read the complete answer and copy the relevant wording into a separate quote column. Keep the ID beside it so someone else can find the source.

Choose quotes that show why a theme matters, not just ones that sound vivid. A useful set may include a typical comment, a comment that shows a different aspect of the issue, and a comment that limits a broad conclusion. If a respondent praises one part of a process while criticising another, do not remove the praise in a way that changes the meaning. Mark any shortened quotation clearly and check your workplace’s rules before using customer wording in a report.

Quotes are evidence of what those respondents said, not proof that everyone felt the same way. Do not ask the assistant to create a quote that captures the general mood. If no response expresses a proposed theme clearly enough to quote, revisit the theme definition or describe the finding without a quotation. Keep the original response and the report excerpt side by side during the final review.

The check against every original response

AI themes become useful findings only after someone checks them against the original responses. Review each proposed theme by opening its cited IDs and asking whether each answer actually supports the definition. Then scan the full response list for relevant answers the AI missed. Check the unclear group as well: a short comment may be meaningful even if the assistant could not place it.

Look for three distinct mistakes. A false match puts a response under a theme it does not support. A missed match leaves a relevant response out. A broad label hides different problems that need different actions. Correct the spreadsheet rather than trying to repair these mistakes only in the written summary. Where a response supports several themes, retain all relevant assignments instead of making the categories artificially exclusive.

Be careful with claims about which issue is most common. Before making that claim, decide whether you are counting responses or theme mentions, since a response can mention several issues. Exclude blanks consistently and check the underlying rows. If the dataset is too large to review fully, check a deliberate mix of responses from each theme and from the unassigned group, then state that the findings were checked on a sample. Do not describe a sample check as a review of every response.

The report and the point where automation helps

A useful survey report names each theme, explains what respondents meant, and shows where the evidence came from. For each theme, include a short definition, the response IDs behind it, a verified quote if appropriate, and a practical question for the team to investigate. Keep uncertain interpretations visible. A complaint about unclear confirmation, for example, suggests checking the booking journey; it does not by itself prove which message caused confusion.

A spreadsheet and an AI chat tool are usually enough for a one-off survey. Save the cleaned response file, theme definitions and checked assignments so a colleague can follow your reasoning. If new responses arrive regularly, reuse the same definitions but review whether new comments fit them. New issues can appear, and an old label can stop being helpful.

Automate collection or routing only when repeated manual work is the real problem. If survey exports arrive as email attachments, automating emails with attachments may help get files into the right place. If the themes reveal recurring service issues, categorising support emails with AI addresses a separate, ongoing inbox task. Neither replaces checking survey findings against the respondents’ original words.

Sources consulted

• Google Forms Help (support.google.com)
• OpenAI Help Center (help.openai.com)

Start a free course