Find recurring support themes without inventing churn

Feedback clustering is useful when it preserves what customers actually said. A small support sample does not establish the cause of churn or the opinion of the whole customer base.

Prepare the input

Supply anonymized feedback with message and customer references where appropriate. Define the period and sampling method. State whether multiple messages may come from the same person. Do not include private details that are irrelevant to the question.

Run the workflow

Ask for themes, supporting references, distinct-customer counts if available, and representative short paraphrases. Keep “asked about cancellation” separate from “cancelled.” Let one message belong to more than one theme when justified.

Prompt to copy

Analyze this feedback sample: [data and period].
Group observed issues into themes with supporting references.
Report message count and distinct-customer count separately where possible.
Separate stated cancellation, cancellation questions, and inferred risk.
Do not invent causes, percentages, or population-wide conclusions.
Suggest one verification question for each material theme.

Review a worked example

Illustrative sample: five messages concern billing, but four are from one customer. The finding is five messages from two customers, not five dissatisfied accounts. A complaint about price is not proof that price caused cancellation. Check every count and the strength of the conclusion.

Use a small manual sample to check theme boundaries. If one theme mixes price confusion with payment failures, split it because the next actions differ. Retain outliers that could matter despite low frequency. A rare but serious access issue should not disappear merely because a more common request dominates the count.

Use it again

Compare later samples using the same definitions and show changes in sampling. A recurring helpdesk feed needs a separate connection and data review. Explore Bot or discuss that integration after the manual analysis proves useful.