ChatGPT Customer Feedback Analysis: From Comments to Actions

Start with one decision

“Summarize our feedback” often produces a polished paragraph that hides the evidence. A more useful question is: which problems should we investigate next, and which comments support each problem?

This workflow turns a small set of comments into a theme table, an evidence trail, and a short action list. It uses text you paste yourself. It does not assume that ChatGPT has access to your help desk or customer database.

Prepare a numbered, anonymized sample

Remove names, email addresses, order numbers, and other details that are unnecessary for the analysis. Follow your organization’s rules for sharing customer information. Keep one comment per row and give each row a stable ID.

The four comments below are an invented teaching example. Their counts describe this sample only.

F01 | I could not find the delivery estimate before paying.
F02 | Please show the delivery date on the product page.
F03 | The sizing guide was hard to understand.
F04 | I was unsure which size to choose.

1. Check the input before grouping

Ask ChatGPT to list duplicate IDs, blank rows, and comments that contain more than one issue. Repeated wording does not prove that two rows are duplicates: different people may report the same problem.

Keep the source text unchanged. If you choose to exclude a row, record the ID and reason so that the denominator can be reconstructed later.

2. Use one primary theme per comment

A primary theme gives you a count that adds up to the number of included comments. Optional secondary tags can overlap, so put them in a separate table.

For the example, a reasonable provisional grouping is:

Primary theme	Count	Supporting IDs
Delivery information	2	F01, F02
Sizing information	2	F03, F04

The total is 4, matching the four input rows. “2 of 4 comments” is valid for this example; “half of all customers are unhappy with delivery” is not supported. Ask for an ID-to-theme mapping when a category is unclear.

3. Separate evidence from the proposed fix

F01 reports difficulty finding delivery information. F02 asks for a particular solution. Neither establishes whether the root cause is a missing date, a poorly placed link, or wording on the checkout screen.

An appropriate next step is to inspect the delivery-information path and test whether people can find it. That is a proposed investigation, not a proven diagnosis. For sizing, review the two comments alongside the guide before deciding to redesign it.

4. Verify one row and then the totals

• Trace each theme back to its listed IDs. Do the quoted or summarized comments actually support it?
• Check that every included ID appears once in the primary mapping and that the theme counts total the included rows.
• Label suggestions as proposals. Leave owners and deadlines unassigned unless you supplied them.
• Keep low-frequency but clearly described issues visible rather than letting a ranking erase them.

Useful follow-up requests

For a larger collection, review manageable batches with the same category definitions, then reconcile the ID mappings before combining counts. Avoid averaging percentages from batches of different sizes.

Show the source IDs behind every conclusion. Flag conclusions that are not directly supported.

If one comment fits two themes, explain your primary assignment and list the secondary theme separately.

Turn the verified findings into three proposed investigations. Do not invent owners, deadlines, refunds, or policies.

Use the shared prompt

The complete reusable prompt is kept in the shared directory, where you can edit the bracketed fields and copy it. The article stays focused on the workflow and verification.

Open and customize the shared prompt

Sources and related reading

Data controls help you manage training preferences; they do not replace your workplace’s sharing rules.

• OpenAI: data controls in ChatGPT
• OpenAI: checking ChatGPT’s accuracy
• Praclane: verify ChatGPT answers

Photo: Kari Shea / Unsplash.