# AI prompts for customer feedback analysis: themes, evidence and priorities you can trust

> How to use ChatGPT, Claude or Gemini to analyze reviews, survey answers and support tickets: define themes before counting, quote the evidence, separate frequency from severity, and avoid invented percentages. Templates included.

Source: Promptea — https://www.promptea.me/en/guides/ai-prompts-for-customer-feedback · Language: English · Last updated: 2026-11-13

## Why feedback summaries go wrong

- Asking 'what are customers saying?' gets a fluent paragraph that mirrors the loudest comments, not the most common ones.
- Models round and estimate when asked for numbers: '40% of users mention pricing' may not come from any count at all.
- Themes drift between runs: 'onboarding', 'setup' and 'getting started' become three themes one day and one the next.
- Sarcasm, mixed reviews ('love it, but…') and non-English comments are easy to misclassify without explicit rules.
- Before: 'Summarize these 300 reviews.' After: 'Tag each review with one theme from this list, quote the sentence that justifies it, and count tags at the end.'

## A workflow that holds up

- Step 1 — discover: on a sample (for example 50 comments), ask for candidate themes with a one-line definition and two example quotes each.
- Step 2 — fix the codebook: merge overlapping themes yourself, add 'other' and 'unclear', and keep that list identical for every batch.
- Step 3 — tag: give each comment an ID and ask for one row per comment (ID, theme, sentiment, quote). Rows are easy to spot-check; paragraphs are not.
- Step 4 — count outside the model: paste the rows into a spreadsheet, or ask for counts only from the rows already produced, never from memory of the input.
- Step 5 — prioritize: rank themes by frequency and by impact (churn risk, blocked tasks, revenue) as two separate columns, so a rare but critical issue isn't buried.

## Privacy and honesty checks

- Remove names, emails, order numbers and other personal data before pasting; keep your own ID so you can trace a row back.
- Check your company's policy on which AI tools may receive customer data.
- Spot-check at least a handful of tagged rows per theme against the original comment before you share the report.
- Report what the data is: 'themes in 312 app-store reviews from October', not 'what our customers think'.

## Templates

### Discover candidate themes from a sample

```text
Below are [number] customer comments from [source: reviews / survey / support tickets] about [product]. Each line starts with an ID.

<comments>
[paste the comments, one per line, as "ID: text"]
</comments>

Propose 6 to 10 themes that cover these comments. For each theme give:
- name (2–4 words)
- definition (one sentence: what belongs and what doesn't)
- two IDs with a short exact quote each

Rules:
- Themes must not overlap; if two are close, merge them and say so.
- Include "other" for comments that fit nowhere.
- Do not count or estimate percentages yet.
```

[Open in Promptea](https://www.promptea.me/en/analyzer?prompt=Below+are+%5Bnumber%5D+customer+comments+from+%5Bsource%3A+reviews+%2F+survey+%2F+support+tickets%5D+about+%5Bproduct%5D.+Each+line+starts+with+an+ID.%0A%0A%3Ccomments%3E%0A%5Bpaste+the+comments%2C+one+per+line%2C+as+%22ID%3A+text%22%5D%0A%3C%2Fcomments%3E%0A%0APropose+6+to+10+themes+that+cover+these+comments.+For+each+theme+give%3A%0A-+name+%282%E2%80%934+words%29%0A-+definition+%28one+sentence%3A+what+belongs+and+what+doesn%27t%29%0A-+two+IDs+with+a+short+exact+quote+each%0A%0ARules%3A%0A-+Themes+must+not+overlap%3B+if+two+are+close%2C+merge+them+and+say+so.%0A-+Include+%22other%22+for+comments+that+fit+nowhere.%0A-+Do+not+count+or+estimate+percentages+yet.&purpose=data&target=claude)

### Tag every comment with a fixed codebook

```text
Classify each customer comment using ONLY the themes below.

<themes>
[paste your final theme list with one-line definitions, including "other" and "unclear"]
</themes>

<comments>
[paste the comments as "ID: text"]
</comments>

Return a table with one row per comment and these columns:
ID | theme | sentiment (positive / negative / mixed) | quote (the exact words that justify the theme)

Rules:
- Exactly one theme per row; if two apply, pick the main complaint or praise.
- If the comment is sarcastic or ambiguous, use "unclear" instead of guessing.
- Copy quotes exactly; do not paraphrase.
- Do not skip any ID and do not add a summary.
```

[Open in Promptea](https://www.promptea.me/en/analyzer?prompt=Classify+each+customer+comment+using+ONLY+the+themes+below.%0A%0A%3Cthemes%3E%0A%5Bpaste+your+final+theme+list+with+one-line+definitions%2C+including+%22other%22+and+%22unclear%22%5D%0A%3C%2Fthemes%3E%0A%0A%3Ccomments%3E%0A%5Bpaste+the+comments+as+%22ID%3A+text%22%5D%0A%3C%2Fcomments%3E%0A%0AReturn+a+table+with+one+row+per+comment+and+these+columns%3A%0AID+%7C+theme+%7C+sentiment+%28positive+%2F+negative+%2F+mixed%29+%7C+quote+%28the+exact+words+that+justify+the+theme%29%0A%0ARules%3A%0A-+Exactly+one+theme+per+row%3B+if+two+apply%2C+pick+the+main+complaint+or+praise.%0A-+If+the+comment+is+sarcastic+or+ambiguous%2C+use+%22unclear%22+instead+of+guessing.%0A-+Copy+quotes+exactly%3B+do+not+paraphrase.%0A-+Do+not+skip+any+ID+and+do+not+add+a+summary.&purpose=data&target=gpt)

### Turn tagged rows into a prioritized report

```text
Write a one-page feedback report for [audience: product team / leadership] from the tagged rows below.

<rows>
[paste the table: ID | theme | sentiment | quote]
</rows>

Structure:
1. Scope: source, period and number of comments (count the rows; do not estimate).
2. Themes table: theme | count | negative count | one representative quote with its ID.
3. Top 3 issues to act on, judged by count AND by impact. Explain the impact in one sentence each and say which part is your judgment.
4. What customers praise (to protect in future changes).
5. Open questions the data cannot answer.

Rules:
- Every number must come from counting the rows above.
- Quote only text that appears in the rows.
```

[Open in Promptea](https://www.promptea.me/en/analyzer?prompt=Write+a+one-page+feedback+report+for+%5Baudience%3A+product+team+%2F+leadership%5D+from+the+tagged+rows+below.%0A%0A%3Crows%3E%0A%5Bpaste+the+table%3A+ID+%7C+theme+%7C+sentiment+%7C+quote%5D%0A%3C%2Frows%3E%0A%0AStructure%3A%0A1.+Scope%3A+source%2C+period+and+number+of+comments+%28count+the+rows%3B+do+not+estimate%29.%0A2.+Themes+table%3A+theme+%7C+count+%7C+negative+count+%7C+one+representative+quote+with+its+ID.%0A3.+Top+3+issues+to+act+on%2C+judged+by+count+AND+by+impact.+Explain+the+impact+in+one+sentence+each+and+say+which+part+is+your+judgment.%0A4.+What+customers+praise+%28to+protect+in+future+changes%29.%0A5.+Open+questions+the+data+cannot+answer.%0A%0ARules%3A%0A-+Every+number+must+come+from+counting+the+rows+above.%0A-+Quote+only+text+that+appears+in+the+rows.&purpose=summarization&target=gemini)

## Frequently asked questions

### Can I paste thousands of comments in one prompt?

Even when the context window allows it, long single runs are harder to check and rows get skipped more easily. Work in batches (a few hundred comments at a time) with the same codebook, then combine the rows in a spreadsheet and count there.

### Should the AI decide which issues matter most?

It can propose a ranking, but frequency is something to count and impact is a business judgment. Ask the model to show both columns and to label its own reasoning, then make the final call with your team.

### How do I check that my tagging prompt is clear enough?

Run the same batch twice and compare the tags; rows that change point at vague theme definitions. Promptea's analyzer also flags missing output-format and constraint instructions before you run the prompt on real data.
