ReviewFold
Guide · A workflow you can audit
Analyze App Reviews with ChatGPT: A Checkable Workflow
A chat assistant will produce a tidy table of themes whether or not it read your reviews carefully. The part you control is the input format, the shape you ask for, and the checking you do afterwards. This page gives you all three, and nothing that requires an account here.
01 · The difference that matters
An answer is useful when you can contradict it.
Two analyses can look identical and differ completely in whether you could catch them being wrong. Three properties separate them.
- Every claim carries IDs
Not “users find the paywall pushy” but the four review IDs that made someone write that. Without them, you are reading a summary of a summary.
- Counter-evidence is requested
If you do not ask for the reviews that disagree with a group, you will get groups with no dissent — which describes a collection process, not a sample.
- “Unclear” is an allowed output
An empty uncertainty column deserves a check: the source text may leave important questions unanswered.
This does not depend on which assistant you use. It depends on what you send and what you accept.
02 · The input
Four fields per review, one line each.
A stable ID is the whole trick. Once each row has a short handle, every claim in the answer can be traced back to one sentence.
Input format · one review per line
Keep the order when you paste
id | rating/5 | date | text
1 | 5 | 2026-08-25 | They let you use alot of stuff for free, you just have to watch an ad in exchange every time you use one and when you export
2 | 1 | 2026-07-26 | Can’t even use without watching ads. When you do watch you don’t get to enhance your photo.Short IDs (1, 2, mapped to the full store review IDs in your source table) are easier to check than long ones, and easier for the model to copy without inventing digits.
What to leave out
- Anything identifying a customer
Support tickets, order numbers, email addresses. Reviews are public; the rest of your inbox is not.
- Your own conclusions
If you paste “the paywall is the main problem” above the reviews, you will get it back as a theme.
- Reviews you have not read
Paste a sample you can defend. Two hundred reviews you scanned once will produce themes you cannot verify.
Pasting reviews into a chat tool sends them to that provider. Check their current data controls before pasting anything beyond public review text — and never paste customer-identifying material to test whether it is handled well.
03 · The prompt
Ask for groups, IDs and disagreement — in that order.
This is an editorial specification, not a recorded session. We did not run it against ChatGPT for this page, so treat it as text you can edit and test yourself.
You are helping me analyse app store reviews. I will paste
reviews between the markers. Each line is:
id | rating/5 | date | text
--- REVIEWS ---
[paste here]
--- END ---
Rules:
1. Use only these reviews. Do not use outside knowledge about
this app, its pricing, or its features.
2. Group reviews by the task the person was trying to complete
and what blocked it. Do not group by adjective or tone.
3. Every group needs: the task in one sentence, at least two
supporting ids, any contradicting ids, and what stays unclear.
4. A group with only one supporting review goes into a separate
list called "single accounts", not into the main table.
5. If two reviews in the same group disagree, say so. Do not
average them into a neutral sentence.
6. Quote wording exactly. Keep spelling and punctuation.
7. If a review is too vague to name a task, put it in "cannot be
assigned" with a reason of ten words or fewer.
8. Do not say whether a problem still exists. You cannot know
that from the text.
Output: first a table with the columns
Group | Task | Supporting ids | Contradicting ids | What stays unclear
then the "single accounts" list, then "cannot be assigned".Two rules do most of the work. Rule 3 asks for traceable evidence. Rule 8 asks the model not to assume a reported problem still exists. Both instructions still need checking.
04 · The output shape
One row, filled by hand, so you can see what “checkable” looks like.
This row was assembled editorially from a separate saved sample of 200 DramaBox reviews (September 2026). It is an illustration of the format, not a model’s output.
| Group | Task | Supporting ids | Contradicting ids | What stays unclear |
|---|---|---|---|---|
| Earning episode access through ads | Continue watching using ads to earn access. | 14578586430 14576703925 |
14577347724 | One reviewer reports too few ad opportunities, another cannot earn coins through ads, and a third describes repeated ads. These are different experiences; the text does not establish why. |
The contradicting column is not a defect. It is the most valuable cell in the row, because it tells you what to go and look at before you write anything down.
05 · Your half of the work
Five checks before using the result.
Run them in order. Each one catches a different way an answer can look right and be wrong.
- Every ID exists, exactly once
Search the paste for each ID in the table. A repeated ID inside one group usually means two sentences were treated as two reviews.
- Separate repeated feedback from single accounts
If a group lists one ID, it belongs in “single accounts”. This check is mechanical and catches the most padding.
- Read each quoted sentence beside its ID
Not beside the group name. The gist is often close and the detail is often wrong.
- Look for negation and condition words
“Until”, “unless”, “except”, “even after”, “not free”. These carry the meaning and are the first thing a paraphrase drops.
- Check that the uncertainty column is not empty
An empty column is a claim of certainty. If the reviews really do settle everything, you should be able to say how.
If a check fails, do not rewrite the prompt from scratch. Paste the failing rows back and ask for that group again with the IDs you know are in scope.
06 · Choosing a route
When the manual route is enough, and when it is not.
We have not compared their accuracy. Either route needs traceable source text and a check of the grouping.
Manual, with a spreadsheet
- Enough when the sample is small, the question is singular, and you will not repeat it next month.
- Enough when you already know the three tasks you are looking for and only need the IDs.
- Keep the source text beside each interpretation so another reader can check it.
- Requires more manual work as the sample grows and similar feedback uses different wording.
A tool that keeps the excerpts
- Worth it when you need the same job done for several apps or several months.
- Worth it when the output has to stay attached to the original reviews for someone else to check.
- Worth it when reviews arrive in more than one language and you need the original beside the reading.
- Not worth it if you would not read the excerpts anyway.
This is the fork the comparison of analysis options covers in more detail, including how collection, saved reports and exports fit into a repeatable workflow.
Boundaries
What this page does not do.
It is text and a button. Everything else happens in the tool you already use.
- No connection to any assistantNo API key, no plugin, no upload. You copy the prompt and paste your own reviews.
- No recorded runThe example row is filled by hand and labelled as such. There is no claimed accuracy, speed or win rate.
- No privacy promiseWhere your pasted text goes is decided by the tool you paste it into, and by its current settings.
- No current-state claimA review describes a moment. Nothing in a chat session can tell you whether the bug is still there.
Current support, stated plainly
- Apple App Store: US and Canada. Google Play: US.
- Up to 200 recent public reviews per report.
- One app and one region per report; no cross-app comparison.
- A free preview shows themes and counts; the paid report adds review excerpts, PDF and CSV.
- No automatic chat integration, no prompt library and no accuracy guarantee.
If you would rather not assemble the sample by hand
Same evidence rules, kept for you.
ReviewFold applies the grouping and leaves the original excerpts attached, so the check is reading rather than re-pasting. You can still overrule any group.