The prompt that turns your reviews into ads

What review mining actually is

This is not an AI trick. It is a copywriting practice, and the person who has been teaching it longest is Joanna Wiebe at Copyhackers. Her line is the whole philosophy: "review mining data should write your copy, you shouldn't write copy". You are not your prospect, so the words that will land on them are sitting in their reviews, not in your head.

What a model adds is speed and consistency at the sorting job. A human mining 40 reviews gets bored around review 12 and starts skimming. The prompt below does all 40 at the same level of attention, counts the themes so you can tell a pattern from a one-off, and refuses to smooth anyone's wording into marketing language on the way through.

Where to get 20 to 50 reviews

1. ⭐ Your own reviews, all ratings. Pull everything, not the five stars. The two and three star reviews are where the hesitations live, and hesitations are what objection-handling copy is made of. If you have a review platform, export the lot. If you do not, copy them by hand.
2. 💌 Your DMs, comments and support emails. Anywhere a customer explained themselves in their own words counts. The question that produced them does not matter, the language does. Strip names and handles before you paste.
3. 🛒 Reviews of a similar product in your category. This is the one that surprises people, and it is the version Copyhackers originally taught, because review mining is aimed at your prospect and the outcome they want, not at your specific product. If you have no reviews yet, mine the category. The problems, the failed alternatives and the anxieties are the same.
4. 🔒 Strip the identifying details first. Names, handles, order numbers, anything that points at a person. You want the language, never the identity, and it saves you a problem later if any of this ends up on screen.

Paste this into Claude

Open a new chat at claude.ai, paste the prompt, fill in the brackets, then paste your reviews underneath. If they live in a spreadsheet, attach the file instead. Claude takes up to 20 files per chat, so an export works as well as a paste.

You are a conversion copywriter who builds ads out of the customer's own words, never your own.

MY INPUTS
- What I sell, the price, and who buys it: [describe]
- Where these came from: [my own reviews, my DMs, my comments, or reviews of a similar product in my category]
- 20 to 50 real reviews, testimonials, DMs or comments, mixed ratings included: [paste]

STEP 1. Mine. Sort exact phrases into 6 buckets: the problem in their life, what they already tried that did not work, what made them start looking, the outcome they actually wanted, the detail they loved, and what almost stopped them buying. Quote word for word and never paraphrase. Mine the negative and middling reviews as hard as the 5-star ones, because that is where the hesitations live. Under each bucket, count how many separate reviews said it and note which review each quote came from.

STEP 2. Rank. Pick the 3 themes that come up most often and carry the most emotion, and use the counts to say why each one beat the others. A theme that is emotional but rare goes in a separate line called the wildcard.

STEP 3. Write. For each of the 3 themes: 3 ad hooks, 1 short ad under 120 words, 1 headline under 12 words, and 1 line that answers the biggest hesitation in that theme. Build every line out of their exact phrases. Where you change a customer's wording, keep their nouns and their verbs, and never upgrade their vocabulary.

STEP 4. Swipe file. List the 15 most quotable customer phrases word for word, and next to each one say where it works hardest: an ad hook, an email subject line, a landing page headline, a caption, or an objection-handling line.

STEP 5. Check. Put the review number next to every line you wrote. Never invent a result, a number, a timeframe or a claim that is not in the reviews. Mark any line that is a claim I could be held to, so I can check it before it runs.

STEP 6. Tell me which buckets came out thin, and give me the exact question to ask my customers to fill each gap.

If I paste fewer than 20 reviews, tell me which buckets are too thin to trust before you write anything.

What each bucket is for

These six are re-voiced from Copyhackers' rapid-fire review mining, and each one feeds a different part of an ad. Knowing which is which is what turns a pile of quotes into copy.

A real run, start to finish

This was run on 33 mixed-rating reviews for an $89 product, including the two and three star ones.

1. Every quote carried its review number. Ten phrases in the problem bucket, eleven in the detail bucket, seven in the hesitation bucket, each one word for word with a number beside it, so any line could be traced back in seconds.
2. Two buckets were flagged as too thin to trust. Four mentions of failed alternatives and five of the trigger moment. Rather than writing from them anyway, it said which ones were thin and why.
3. It wrote the question to fix each gap. For the thin trigger bucket: "what finally made you go and look for something, was there a specific moment, conversation or bad day that pushed you". That is a survey question you can send today.
4. A wildcard theme, kept separate. One theme was rare but the most emotionally loaded in the set, so it was flagged as a retargeting or email angle rather than promoted into the top three on feeling alone.
5. Every risky line was marked. Anything with a timeframe, a number, a health-adjacent statement or a price comparison came back flagged, with the reason. A customer saying a competitor cost four times more is their comparison, not a price claim you can run in your own voice.

4 fixes that almost always work

The landing page prompt

The swipe file is worth more than the ads, because the same phrases work everywhere. Run this in the same chat once the mining is done.

Using only the customer language you have already mined, write the copy for a landing page for this product.

Give me the headline, the subheadline, three benefit sections with a subheading and two sentences each, an objection-handling section written as short questions and answers, and the copy that sits directly above the buy button. Take the headline from the theme you ranked first, take each objection from the bucket of what almost stopped them buying, and build the benefit sections from the outcome they actually wanted rather than from features. Keep every phrase traceable, so put the review number beside every line exactly as you did before. Where a section needs a claim I have not earned from the reviews, leave a clearly marked gap with a note about what evidence would fill it, rather than writing something plausible. Finish by telling me which three customer phrases you did not use and where they would work better.

Plans, privacy and the ChatGPT version

1. 🆓 The free plan handles this. Forty reviews is a large paste but well within a normal chat. A paid plan matters if you want the mining, the writing and the landing page prompt in one continuous conversation.
2. 🔒 Strip names before you paste, always. You want the language, never the person. It is also what keeps you safe if any of this ends up in a screenshot, a deck or a video.
3. ⚖️ A customer's claim is not your claim. A review saying something fixed their back is their experience. Repeating it in your brand voice turns it into a claim you can be held to, which is why Step 5 flags it. Keep those lines as attributed testimonials, not as headlines you assert. It is the same reason ad strategists separate the hook from the claim.
4. 💬 It runs in ChatGPT unchanged. Paste it as it is. The quoting discipline and the counts both hold up, so the only real difference is formatting.

Proof

Joanna Wiebe on review mining: the data should write your copy, you shouldn't
Copyhackers on rapid-fire review mining and the categories the six buckets come from
Why a claim is the part of an ad you can be held to
Anthropic on grounding answers in quotes and asking a model to check its own work
What you can upload to Claude, and the 20 files per chat limit