Voice of Customer Research With AI: Mine Reviews and Calls for Copy

A copywriter can mine customer reviews and call transcripts with AI by gathering the raw text, removing personal details, asking for coded themes with a source ID on every quote, counting the themes in a spreadsheet, then checking each phrase against the original. The output is ranked themes, exact customer phrases and a swipe file you can write from.

The value is in the exact wording. Customers describe their problem in sentences that no brief contains: "I kept putting it off because I didn't want to look stupid asking", "it just takes the dread out of the end of the month". Those lines are headlines in waiting. AI is good at sorting a few hundred of them quickly and very good at fabricating a quote that sounds right, so the method below is built around proof of origin.

Follow me on Instagram@sagnikteaches

What counts as voice-of-customer data, and where to find it

Voice of customer research means collecting how buyers describe their situation, the outcome they wanted and the words they use, in their language, not yours. Good sources, roughly in order of richness:

Connect on LinkedInSagnik Bhattacharya

• Sales and discovery call transcripts. The customer explains the problem in their own words before they know the "right" terms.
• Onboarding and support emails. The first thing people ask is usually what they feared.
• Reviews of the client's business and of competitors, on public review sites and app stores. Three-star reviews are the most honest.
• Survey free-text answers. "What nearly stopped you from buying?" is the best single question.
• Community threads and comments where your client's buyers talk to each other.

Say a copywriter has been hired to rewrite the website of an accountancy practice that wants more small-business clients. The client shares 6 call transcripts from new-client calls, 94 reviews of the practice, and 70 reviews of two competitors. That is 164 reviews plus 6 calls, which is plenty for a first pass.

Subscribe on YouTube@sagnikteaches

Permission and privacy before you upload anything

Get this settled first, because it changes where the files can go.

1. The client's consent. Transcripts and emails belong to your client's relationships. Confirm in writing that you may analyse them, and for what.
2. The customer's knowledge. Recorded calls should have been disclosed as recorded. If you are not sure, ask the client what customers were told, and if the answer is nothing, don't use the recordings.
3. The tool. Use a business plan that does not train on content by default (ChatGPT Business, Claude Team and Gemini in Workspace work that way per vendor pages), or an individual plan with training switched off and with redacted text only. Two-seat minimums apply on the team plans of ChatGPT and Claude, so a solo freelancer often uses an individual plan.
4. Public reviews. Reading and analysing public reviews for research is common practice, but check the review site's terms before you export in bulk; many forbid scraping. Copy a sample by hand if in doubt.
5. Redaction. Replace names, company names, emails, phone numbers and addresses with placeholders. Keep roles ("owner of a four-person plumbing firm") because they matter for segmenting.

A short routine for the redaction is in how to anonymise client data before you paste it into AI. Keep an untouched copy of each source file in a folder only you control, because you'll need it to verify quotes later.

Prepare the corpus so every line has an ID

Give every piece of text a unique ID: R001 to R164 for reviews (R095 onward for competitors), T1 to T6 for transcripts, and S01 upward for survey answers. For transcripts, split by speaker and number each customer turn (T3-14 is customer turn 14 in transcript three). Keep the speaker labels, because you want the customer's words, not the salesperson's.

Put reviews in a spreadsheet with columns: ID, source, star rating, date, text, segment tag. Splitting into batches of 40 to 60 reviews keeps the model accurate; with 164 reviews, run four batches of about 41 and combine afterwards. Long pasted lists make models skip or merge items, so keep each run small enough that you can check the numbering.

The extraction prompt: themes, phrases and evidence

Ask for structured output with fixed fields, not a summary. A summary loses the exact words you came for.

You are helping a copywriter with voice-of-customer research for an accountancy practice that serves small businesses.
Below are customer reviews, each with an ID. For each review extract, as JSON-style rows:
ID | situation before (what was going wrong) | desired outcome | emotion words used | objection or fear | exact phrase worth reusing (copy it character for character) | star rating.
Rules: the exact phrase must appear in the text; do not correct spelling or tidy it. If nothing is quotable, write NONE.
Do not infer anything that isn't written. Do not merge reviews.

REVIEWS:
[R001] ...
[R002] ...

Illustrative output for two reviews: "R014 | leaving bookkeeping until the last weekend of the month | wanted to stop dreading the deadline | dread, stress | worried they'd be judged for messy records | 'it just takes the dread out of the end of the month' | 5" and "R031 | switched after being passed between three people | wanted one person who knew the business | frustrated, fobbed off | fear of explaining everything again | 'I was sick of explaining my business to someone new every time' | 4". Both phrases are usable ideas for copy, and both can be proved with a search for the string in the source.

Run the phrase check on every quote

Models sometimes lightly improve a quote. Use your spreadsheet's search to confirm each "exact phrase" exists in the matching source ID. In Excel or Google Sheets, a formula such as =ISNUMBER(FIND(phrase_cell, text_cell)) returns TRUE when the phrase is present exactly, so you can tick down a column in seconds. Use FIND rather than SEARCH here: SEARCH ignores capital letters (and in Excel treats ? and * as wildcards), so a quote the model has quietly altered can still pass, while FIND matches case and takes the characters literally. Any FALSE gets fixed against the original or deleted. In an illustrative run of 60 extracted quotes, expect a handful to fail on punctuation or a swapped word; those are the ones that would have embarrassed you if they went into a testimonial.

From rows to ranked themes

Once you have rows, ask the model to cluster them, then count in the spreadsheet. Do not trust the model's own counts ("most customers mention..."), because it estimates; counts come from the sheet.

Here is a table of extracted rows. Propose 8 to 12 themes that group the "situation before", "desired outcome" and "objection" columns.
For each theme give: a short name, a one-line definition, and the list of IDs that belong to it.
An ID may belong to more than one theme. List unclassified IDs separately. Do not give counts or percentages.

Then count IDs per theme with COUNTIF or a pivot table, and add average star rating by theme. A simple layout of illustrative results for the accountancy example:

Those numbers are made up to show the shape. In a real run, the most frequent theme is not always the best headline; a smaller theme with strong emotion ("fear of looking foolish") can be the one that separates your client from competitors. Look at which themes appear only in your client's reviews versus competitors', because that is where positioning lives. For a similar approach on a restaurant's review set, see what diners are really saying: using AI to analyse your reviews.

Mining calls: the pattern in how people explain the problem

Calls carry something reviews don't: the order in which people explain their situation, and the words they use before they've been taught yours. After redaction, ask the model to pull the customer's turns only and tag each with what they are doing: describing the problem, describing a failed attempt, naming a trigger ("my last accountant retired"), stating a constraint (budget, time), or raising an objection.

Two prompts matter most:

From the customer turns below, list every sentence where the customer describes what happened just before they decided to look for help.
Return: turn ID | exact sentence | what this trigger tells me.
Then list every sentence where they describe something they tried that did not work. Same format.
Quote only. Do not paraphrase.

Triggers are gold for copy: "my last accountant retired", "I got a letter I didn't understand", "we hired our first employee". Each one is a moment your client's ad or landing page can speak to. A failed attempt is an objection to answer: "I tried doing it myself on a spreadsheet and gave up in March." Illustrative finding from the six calls: four of six customers mentioned a letter or notice they didn't understand as the trigger. That's a headline idea ("Got a letter you can't make sense of?") that no reading of the client's existing website would have produced.

The swipe file: where the exact words live

A swipe file here is a spreadsheet of customer language, not a folder of other people's ads. Columns that earn their place:

1. Phrase (exact, in quotation marks)
2. Source ID and source type (review, call, survey)
3. Theme
4. Emotion (relief, dread, embarrassment, pride)
5. Where it could go (headline, subhead, bullet, objection handler, email subject line)
6. Permission status (public review, consented, paraphrase only)
7. Used where, and what happened

The permission column keeps you honest: a phrase from a public review can inspire a headline in your own words, but placing a named person's quotation on the page as a testimonial requires their consent. If the rules around AI and testimonials are unclear to you, where the legal line is on AI-written testimonials explains it, and the practical sign-off steps are in what a personal trainer can and can't edit in a testimonial.

Turning research into copy: three worked moves

Headline from a trigger. Research finding: four of six calls begin with a confusing letter. Draft headline: "Got an official letter you can't make sense of? Send it to us and we'll tell you what it means." Check: does it use the customer's words ("make sense of") rather than yours ("correspondence from authorities")? It does.

Objection handler from a fear. Finding: 14 reviews mention fear of looking foolish. Draft: "Nobody here will tell you that's a silly question. Most of our clients start by saying they're not sure they're doing it right." Test: it reflects the phrase "I didn't want to look stupid asking" without quoting it, so there's no permission issue.

Subject line from relief. Finding: "dread" appears in 27 reviews. Draft subject: "Take the dread out of month-end". The paraphrase keeps the emotion without lifting a named customer's sentence.

For each draft, ask the AI to write five variants only after you've shown it the swipe-file lines, and tell it to use no word that isn't in the customer's vocabulary unless needed. Then choose and edit by hand. Best AI for copywriting ads: variants, testing and client sign-off shows how to test several.

Competitor and category reviews: finding the gap

Run the same extraction on competitor reviews. Compare themes side by side: where are competitors praised for something your client doesn't mention? Where do competitors' negative reviews (one and two stars) complain about something your client does well? In the accountancy example, 22 of 70 competitor reviews complain about not being able to reach a person, while only 3 of 94 of your client's reviews do. That contrast is a promise your client can make with evidence behind it.

Be careful about reading too much into small samples. A theme from 3 reviews is a hypothesis; confirm it with a question on the next survey or call.

Transcript clean-up: speech-to-text errors that distort the words

Transcripts from call recorders contain mistakes that matter when you are hunting for exact phrases. Names of products get misheard, filler words pile up, and sentences split across speaker turns. Before extraction, do a quick pass: ask the AI to list every word that looks like a transcription error or an unclear term, without correcting anything, and fix those yourself from the audio if you have it. Illustrative flags: "payroll" heard as "pearl role", "invoice" heard as "in voice". Remove filler ("um", "you know") only from the swipe-file copy of a phrase and mark it as lightly edited, never from the source. If the phrase will appear anywhere as a quotation, listen to the audio for that sentence and confirm what was said.

Better questions for the next round of research

The best VoC data comes from a few well-chosen questions. If the client can email a short survey, these five produce language you can use, and the AI can then code the answers with the method above:

1. What was happening just before you decided to look for help with this?
2. What had you tried before, and why did it not work?
3. What nearly stopped you from choosing us?
4. How would you describe what we do to a friend in one sentence?
5. What is different now?

The fourth question is the most useful for copy, because it hands you the customer's own category word. If six customers in ten say "bookkeeper" while the website says "financial wellbeing partner", you know which to use. In an illustrative result from 40 answers, 23 used some form of "someone who sorts out my accounts", 9 said "tax person" and 3 said "adviser"; the headline picks the first phrase.

Segment the data before you write

One corpus often hides two audiences. In the accountancy example, tagging each row as "new sole trader" or "established firm with staff" splits the 94 reviews into 38 and 56. Sole traders talk about fear and confusion ("I don't understand the forms"); established firms talk about time and reliability ("I need it done without chasing"). Run the theme prompt separately on each segment and compare. Illustrative result: "being passed between people" is 6 of 38 for sole traders but 13 of 56 for established firms (the 19 in the theme table), so the continuity promise belongs mainly on the page for firms with staff. Without the split, the theme looked equally strong everywhere and the page would have spoken to nobody in particular.

Ask the model to tag the segment from role cues in the text only ("my staff", "my employees", "I'm just me"), and check a sample of ten tags yourself. Anything ambiguous goes into an "unknown" bucket rather than being guessed.

A message map you can hand to a client

When the themes are ranked, put them on one page the client can approve. Illustrative map for the accountancy practice:

Mark every cell with the source IDs behind it. A client who asks "where did you get that?" gets an answer in seconds, and that is what makes the research credible rather than a creative guess.

Mistakes that make VoC research worthless

• Asking for a summary. A summary is the model's language. You want the customer's.
• Letting the model count. Counts must come from the sheet; models guess.
• Only five-star reviews. They flatter. Three-star and one-star reviews hold the objections.
• Mixing segments. A homeowner and a business owner use different words. Tag each row and analyse segments separately when you have enough data.
• Using phrases from the client's own marketing. If a "customer" quote echoes the company's slogan, it may have been prompted by it. Flag those and weigh them less.
• Skipping the check that the phrase exists. One invented quote in a client deck costs you more than all the time you saved.

Three tests before the research reaches a client

Run three tests before you hand anything over. First, the source test: pick ten random quotes from the swipe file and find each in the originals; all ten must match. Second, the stranger test: show your top three themes to the client's sales lead, who spoke to many customers, and ask whether any surprises them or whether any is missing. Third, the copy test: put two headlines, one in the old wording and one in customer language, in front of a small email list or a short ad test and see which gets more clicks. The aim is evidence that the new language lands better, not proof that AI did a good job.

For the broader picture of where AI sits in a copywriter's workflow, ChatGPT prompts for copywriters: research, angles and edits goes through it. Time check on the example: about 3 hours to redact and ID 164 reviews and 6 transcripts, 2 hours for extraction runs and quote checks, 1.5 hours to count and tidy themes, and 2 hours to write the first batch of copy, so about 8.5 hours against the several days a manual read would take. If you'd rather map the process for your own client work, that's what my AI implementation consultation covers; you can contact me here.

Questions copywriters ask about mining customer language

How many reviews or calls do I need?

Around 50 to 100 reviews or five to eight calls is usually enough to see the main themes repeat. If a theme only appears once, treat it as a lead for a question, not a finding. Keep going until new material stops producing new themes.

Can I quote customer phrases in copy?

Short phrases lifted from public reviews are commonly used as inspiration, but exact quotes shown as testimonials need the person's permission. Paraphrase in your own words for headlines, and ask for written consent before using anyone's words, name or story as a named testimonial.

Will AI invent quotes?

Yes, sometimes. It can tidy a phrase and present it as verbatim. Always require a source ID next to every quote, then search your original file for the exact string before it goes into a swipe file or a client deck.

Is it safe to upload call recordings or transcripts?

Only if the people were told about the recording and its use, the client has agreed, and you use a plan that does not train on your content. Remove names, numbers and any detail that identifies a person before analysis.

Further reads

• Can AI Help Copywriters? Research, Drafts and Edits for Client Work — Where AI fits across research, drafts and edits.
• How to Review AI Call Transcripts for Quality and Compliance — Reviewing call transcripts for quality and compliance.
• How to Write Case Studies From Customer Interviews Using AI — Turn customer interviews into case studies.
• AI Software for Copywriting: A Freelancer's Monthly Tool Budget — What a freelancer's AI tool budget looks like.
• Shadow AI: How to Stop Staff Pasting Client Data Into Free Tools — Stop pasting client data into free tools.
• Reusing Past Client Work With AI Without Leaking Client Data — Reusing past client work without leaking details.
• AI Tools for Copywriting: A Freelancer's Draft-to-Delivery Workflow — One landing-page job walked from brief to delivery: client pack, AI draft, human edit, claims log, plagiarism check and an honest AI-use line.
• Best AI Model for Copywriting: How to Test ChatGPT, Claude and Gemini on One Brief — An owner's brief-test scorecard: one brief, one voice sample and one edit round through ChatGPT, Claude and Gemini, and where each needs heavy editing.
• Best AI for Writing Copy: A Freelance Copywriter's Edit Workflow — A four-stage workflow for freelancers: voice sample, AI draft, line-by-line human edit and a plain disclosure note to the client.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI ChatGPT Business pricing page and Claude pricing page (seat minimums, no training by default); Google Workspace generative AI privacy hub; Microsoft Support pages for the FIND and SEARCH functions; Google Sheets FIND help; standard practice for coding qualitative customer data.