How to mine reviews for customer language with AI

This guide is part of our series on AI for branding and design.

Businesses describe themselves in their own vocabulary: “comprehensive solutions,” “full-service,” “client-focused.” Customers describe their problem in theirs: “the AC died during my kid’s birthday party,” “I was tired of chasing my old agency for reports.” The second kind of language makes better headlines and ads, because it is what the reader is already thinking.

Review mining collects that language at scale. AI reads hundreds of reviews and threads and pulls out the exact phrases people use for the problem, the moment they went looking and the outcome they wanted.

What you need

• An AI assistant that accepts file uploads, such as Claude
• The client’s reviews from Google, industry sites and its own testimonial pages
• Competitors’ reviews from the same places
• Reddit threads and forum posts in the niche, found by searching the problem in plain words
• A spreadsheet to hold the phrases

Step 1: Collect the text

Copy reviews into a spreadsheet with columns for source, business, star rating, date and review text. Many review platforms limit automated scraping in their terms, so copying by hand, using the business’s own review export, or using a listing tool you already pay for is the safer route. For Reddit, search the problem, not the industry: “contractor never showed up” finds more honest language than “home services.”

Aim for 200 or more reviews across the client and three to five competitors. Include the one, two and three star reviews. They hold the fears and frustrations that the best ads answer.

Step 2: Pull out the exact phrases

Attached is a spreadsheet of reviews and forum posts about [category].
Extract exact phrases (3 to 15 words, copied word for word) in these groups:
1. How people describe the problem before they bought
2. What made them start looking now (the trigger)
3. What they were afraid of or burned by before
4. The outcome they wanted, in their words
5. What surprised them, good or bad
For each phrase, give the source row and the star rating. Do not paraphrase
or tidy the wording. Then count how many times each idea appears across the
file, even when the wording differs.

The “do not paraphrase” line is the important one. AI likes to clean up language, and the clean version is the one you are trying to get away from.

Step 3: Compare with how the business talks

Give the AI the client’s homepage and service page copy alongside the phrase list:

Compare the attached website copy with the customer phrase list. List:
1. Ideas customers mention often that the site never mentions
2. Words the site uses that no customer uses
3. Customer fears the site does not answer
Quote both sides.

This table is usually the most persuasive thing you can put in front of a client. It shows, in their customers’ words, what their site is missing.

Step 4: Turn phrases into copy

Ask for headline and ad options built only from the phrase list, each labeled with the phrases it uses. Keep the customer’s wording as close to intact as you can. Then a writer picks and edits. The best lines often need almost no change.

Checks before you use them

• Do not quote a review in an ad or on a page without permission, and never present competitors’ reviews as your own customers’ words. Use competitor reviews to learn the language, not to copy it.
• Check that the frequent ideas are frequent in the raw data, by searching the spreadsheet for a few of them.
• Refresh the list once a year. Customer language shifts with prices, seasons and news.

These phrases feed the word list in your brand voice guide and make the strongest options for persona message testing.