If you sell a product, the cheapest research you can get is what customers already wrote about the competing one. A model can read 150 reviews in a minute and hand back themes. It can also invent themes, miscount, and flatter whatever you already believe. The prompts below are built to limit that.
Where can I legally and sensibly get reviews to analyse?
Start with the source, because it decides whether the whole exercise is fine.
- Your own reviews from your store, marketplace seller account and support tickets. Lowest risk.
- Exports your seller or store platform gives you for your own listings.
- A small hand-copied sample of public competitor reviews, for private analysis. Slow, but you read them while you copy, which is useful.
- Not recommended: automated scraping of marketplaces or review sites. Their terms often restrict it, and I haven't verified any one site's current terms in this session, so read them yourself. Our post on AI web scraping tools covers the technical side, but technical possibility isn't permission.
Strip names, usernames, order IDs, phone numbers and any personal detail before pasting. See what not to paste into AI tools.
What is the right way to prepare the reviews?
One review per row in a spreadsheet, with these columns: id (a number you assign), stars, date, text, and source (your store, competitor A, competitor B). Number them yourself. The id lets you verify any claim the model makes by looking up the exact review.
Keep each competitor in a separate batch. Mixing them invites the model to blur one product's problems into another's.
Prompt 1: tag every review before summarising anything
Tagging first, summarising second, is the main trick. It forces the model to cite evidence.
You are analysing customer reviews for [product type].
INPUT: a numbered list of reviews. Each line: id | stars | text.
TASK: For each review, return one row with:
- id
- sentiment: positive / mixed / negative
- features_praised: short phrases quoted or closely paraphrased from the review (max 3)
- complaints: short phrases (max 3)
- unmet_wish: something the reviewer wanted that the product lacks, or empty
- use_case: what they were using it for, or empty
RULES:
- Use only what the review says. If a field is not stated, leave it empty.
- Do not merge or summarise across reviews in this step.
- Output a table with one row per review. Do not skip any id.
REVIEWS:
[paste batch of up to 40]
Run batches of about 40, not 400. Long inputs are where models skip rows. Afterwards, check that the row count equals the review count. If rows are missing, rerun those ids.
Prompt 2: group the tags into themes you can count
Paste the tagged table into a spreadsheet and normalise the wording yourself, or let the model propose a theme list that you review:
Here is a tagged table of [N] reviews.
Propose a theme list for complaints and for features praised (at most 12 themes each). For each theme give: name, one-line definition, and the ids of the reviews that belong to it.
Rules: a review can belong to more than one theme. Do not output counts; I will count. Mark any theme with fewer than 3 ids as "weak". Do not invent ids.
Then count in the spreadsheet with COUNTIF or a pivot table. Counts you computed from ids you can look up are findings. Counts the model asserts are guesses.
Prompt 3: feature gaps and objections
This is where the research turns into decisions. Give it your own product facts so it compares against something real.
MY PRODUCT FACTS: [features, price band, materials, warranty, known weaknesses]
COMPETITOR THEMES (counted by me):
[paste theme | complaint count | praise count | example review ids]
TASK:
1. Gaps: complaints about competitors that my product already addresses. For each, say what fact of mine addresses it and how confident that match is.
2. Risks: praised features competitors have that mine lacks.
3. Objections: reasons a buyer might hesitate on a product like this (from complaints and unmet wishes), each with a one-sentence answer I can truthfully give from my facts, or "I have no answer".
4. Questions to test: 5 things I should verify (with my own listing, a poll or customer calls) before acting on this.
Do not recommend claims I cannot back up. Do not name competitor brands in any suggested copy.
The "I have no answer" clause is important. A model otherwise cheerfully invents reassurance.
Illustrative example (hand-written, not a real model run)
To show the shape of the output, here is an invented mini case for a hypothetical insulated bottle. The numbers are made up for illustration and are not from any real product or study.
| Theme | Complaint mentions (from a pretend 60 reviews) | Example ids |
|---|---|---|
| Lid leaks when bag is tilted | 11 | 4, 9, 17, 22 |
| Paint chips after a few weeks | 7 | 6, 31, 40 |
| Too heavy to carry | 5 | 12, 28 |
A good Prompt 3 response on this would say: "Leaking lid is the biggest theme. Your facts include a silicone-gasket lid, so this is a possible gap. Confidence: medium, because 'gasket' isn't the same as leak-tested. Test: run a leak test on your own samples and say what you did before claiming it." That last step, doing the real test, is the point.
What biases should I expect?
- Selection bias. People who bother to review are not typical buyers.
- Recency and incentive. Some reviews are incentivised, some are fake. Marketplaces try to filter these, but you can't tell from the text alone.
- Model bias. The model tends to over-weight vivid complaints and produce neat, round themes. Spot-check by reading 10 random reviews against their tags.
- Product mismatch. A complaint about a competitor's price band may not apply to your price.
- Survivorship. You see reviews for products still listed. Dead products with the same flaw are invisible.
A simple check: sample 10 tagged rows at random, compare tag to review text, and note the error rate. If more than 2 or 3 are wrong, fix the prompt before you trust any theme.
How do I turn findings into listing and ad copy safely?
Findings go into your own words, based on your own facts. Never copy competitor review text into your listing, and don't quote reviewers. If you want copy from the themes, use the product description prompts, and for ad angles see the ad creative prompts for ecommerce. If you want more of your own reviews to mine, the post-purchase flow in the email and SMS prompts includes a review-request message that doesn't offer anything in exchange. If reviews point to support issues, the support prompts for returns and FAQs are the next step. For a related Indian-market view of how far to automate a store, see AI ecommerce automation for Indian Shopify sellers.
Checklist
- Source is your own data or a small hand-copied sample, with terms read.
- Personal details removed; each review has an id.
- Tag first, theme second, count in a spreadsheet.
- Each competitor analysed separately.
- Ten random rows spot-checked for tagging errors.
- Every action has a test before you commit money to it.
