Mining Customer Interviews and Reviews With Claude

#01Why is review mining worth your time?

Because customers describe their problems differently from how marketers do. A team might say "streamlined approval workflows" while customers say "I'm tired of chasing my manager for sign-off". The second phrase is what people type into search engines and AI assistants, and it's what makes an ad or landing page feel like it was written for them.

Reading 500 reviews by hand takes days, so most teams skim twenty and draw conclusions. Claude can read all of them consistently and point you to the comments that matter. You still decide what it means.

Our view: review mining is one of the highest-value uses of Claude for marketers, because the output, real customer language, improves almost every other piece of work in this track.

#02What sources can you mine?

Anything where customers write in their own words:

• Reviews on your own site, app stores and review platforms
• Support tickets and live chat transcripts
• Sales call notes and win-loss interviews
• Open-ended survey answers, including NPS comments
• Community forum threads and social replies
• Competitors' public reviews, which show what buyers want from the category

Export to CSV or Excel where possible. CSV is among the file types Anthropic lists for upload (Anthropic), and with code execution, which is available on every plan, Claude can process CSV files by writing and running analysis code (Anthropic). One comment per row, with columns for date, rating and source, works best.

#03How do you handle personal data first?

Before you upload, remove names, email addresses, phone numbers, order numbers and anything else that identifies a person. Keep only the comment text and the fields you need for analysis. Check what your plan and your company policy allow. On consumer plans, Anthropic lets you choose whether your chats can be used to improve its models; on Team and Enterprise plans, commercial terms apply. If in doubt, ask your data protection lead before you start.

Claude can help with the clean-up, but do it in a way you can check. A reliable approach is to delete identifying columns in your spreadsheet tool first, then ask Claude to scan the remaining text for any names or contact details it can find and list the row numbers for you to fix.

#04How do you agree a coding frame?

A coding frame is the list of themes you'll tag each comment with. Let Claude propose one from a sample, then fix it before tagging everything. Changing the frame halfway through ruins comparisons.

<comments>[upload CSV]</comments>

Read a random sample of 150 comments. Propose a coding frame
of 10–15 themes for what customers talk about.
For each theme: a short name, a one-line definition, and two
example comments (exact text, row number).
Separate themes about problems, outcomes, features and buying
decisions. Include an "Other" theme.

Review the frame. Merge themes that overlap, split themes that are too broad, and rename them in your team's language. Ask Claude to check that each definition is clear enough that two people would tag the same comment the same way.

#05How do you tag every comment?

With the frame agreed, ask Claude to tag the whole file. For files of more than a few hundred rows, ask it to use code execution so the results come back as a file you can open in a spreadsheet.

Using this coding frame [paste], tag every comment.
A comment can have up to three themes. Add a sentiment column
(positive, negative, mixed) and a "quotable" column (yes/no)
for comments that state a problem or outcome vividly in one or
two sentences.
Return a CSV with row number, original text, themes, sentiment
and quotable. Then give counts by theme and sentiment.

Then check. Pick twenty random rows and tag them yourself. If you disagree with Claude on more than a few, refine the definitions and run it again. This step is what makes the results trustworthy enough to share.

#06How do you turn themes into messaging?

Counts tell you what's common; quotes tell you how to say it. Ask Claude to build a message map from the quotable comments.

Using the tagged file, for the 5 most frequent problem themes:
1. Summarise the problem in one sentence in customer language.
2. List the 5 most vivid exact quotes (row numbers).
3. List the words and phrases customers repeat most.
4. Suggest one headline and one subheadline that use those words.
Only use phrases that appear in the comments.

The rule to use only phrases from the comments keeps the result honest. Anthropic's prompting guidance recommends asking Claude to extract quotes before analysing, to ground its answers in the source material (Anthropic).

#07What else can review mining tell you?

Several things marketers often miss:

• Buying triggers. Ask Claude to find comments that mention what was happening when the customer started looking. "We hired our fifth employee" or "our auditor complained" are triggers you can target.
• Alternatives considered. Comments often mention the spreadsheet, agency or competitor the customer used before. That tells you who you really compete with.
• Objections that were overcome. "I was worried it would be complicated, but…" is a ready-made objection and answer for your landing page.
• Expectations versus reality. Negative reviews about missing features often reveal a promise in your marketing that the product doesn't keep.

Ask for each of these as a separate pass, with quotes, so the analysis stays focused.

#08How do you mine competitors' reviews?

The same way, using public reviews. Ask Claude to compare the themes in your reviews with those in a competitor's, and highlight problems customers mention about them that customers rarely mention about you. Those are strong angles for comparison content, as long as you describe them fairly and can support them.

Stick to publicly available reviews, follow each site's terms, and never paste in a competitor's private material.

#09How do you keep the insight alive?

Save the coding frame, message map and top quotes to your brand-voice project from lesson 4. Then every draft Claude writes can draw on real customer language. Repeat the analysis each quarter with the same frame, so you can see which themes are growing. A scheduled task on a paid plan can produce a monthly summary from a folder of new exports.

#10From MentionX data: how customer questions shape AI answersOriginal data

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.

The phrases customers use are also the prompts they type into AI assistants. MentionX's aggregates show how answers and citations vary by intent across the prompts we track:

Our view: use the problem themes from your review mining as the starting point for your AI prompt set. Our guide to building a prompt set explains how many you need.

Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.

#11Worked example: Ledgerline listens to 800 reviews

Ledgerline is a fictional accounts-payable tool. Figures are illustrative.

Ledgerline exports 800 app-store and review-site comments, deletes the name and email columns, and uploads the file to Claude. The proposed frame has fourteen themes; the team merges two and renames three.

Claude tags all 800. In a twenty-row spot check, the team disagrees on three, so they tighten the definition of "approval delays" and rerun it. The biggest problem theme is chasing sign-off from managers, and the most repeated phrase is "stuck waiting".

The new homepage subheadline becomes "Stop getting stuck waiting for sign-off." A buying-trigger pass also finds that many customers started looking after a missed supplier payment, which becomes a new ad angle. Ledgerline and these details are fictional.

#12How do you present the findings to your team?

Lead with quotes, not counts. A slide with five real customer sentences about a problem persuades a product or sales team far more than a bar chart. Ask Claude for a one-page summary with the top themes, the share of comments in each, three quotes per theme and one recommended action per theme. Then add the counts as a backup table for anyone who wants the detail.

Be honest about the limits. Reviews over-represent very happy and very unhappy customers, and support tickets only show people who had a problem. Say which sources you used and roughly how many comments, so nobody mistakes the findings for a survey of all customers.

#13What should you try before the next lesson?

Export one source of customer comments, even 100 rows, and run Steps 1 to 3. Compare the message map with your current homepage copy. If the words don't match, you've found your next test.

Next: editing and fact-checking with Claude. Or return to the Claude for Marketers track.

#14How do you mine sales calls and interviews?

Transcripts from sales calls and customer interviews are richer than reviews, because people explain their reasoning at length. They're also more sensitive. Get the right consent for recordings, remove names and company details before uploading, and keep the files in a project that only the right people can see.

The approach changes slightly. Instead of tagging short comments, ask Claude to summarise each transcript into a fixed structure first: the problem in the customer's words, what triggered the search, alternatives considered, objections raised, and what made them decide. Run that for every transcript, one chat or one batch at a time, and save the summaries. Then analyse the summaries together, just as you would analyse reviews.

Win-loss interviews deserve special care. Ask Claude to separate reasons the customer gave directly from reasons it infers, and to quote the exact words for each. Sales and product teams will want to see the evidence, and inferred reasons are easy to over-read.

#15How do survey answers differ?

Open-ended survey answers are usually short and answer a specific question, so the coding frame should reflect that question. If you asked "What nearly stopped you from buying?", the themes will be objections, and the frame should be built around them. Keep the question text in your prompt so Claude interprets answers in context.

Surveys also let you compare groups. If your export includes plan type, company size or region, ask Claude to compare theme frequencies between groups. You may find that smaller customers talk mostly about price and larger ones about permissions, which is valuable for segment-specific messaging. Treat small groups with caution; a theme mentioned by four people isn't a trend.

#16How does this connect to the rest of the track?

The message map feeds directly into later lessons. Lesson 16 uses it for ad copy variants, lesson 17 for lifecycle emails and lesson 18 for social posts. The problem themes also become the starting point for checking how AI assistants answer your customers' real questions, which lesson 19 covers. If you only have time for one piece of analysis this quarter, make it this one, because so much of the rest depends on it.

#17What does a coding frame look like in practice?

Agree the frame before tagging. Ask Claude to propose one from a sample, then edit it yourself and save it as a file so every batch uses the same codes.

Here are 60 reviews (a random sample). Propose a coding frame:
8-12 themes, each with a one-line definition, an example quote,
and what does NOT belong in it. Include a theme for "other"
and one for praise that names a specific feature.
Then list themes you considered but merged, and why.

Then tag the full set and output a table you can sort:

Using coding_frame.md, tag each review below. Output a table:
review_id | themes (max 2) | sentiment (pos/neg/mixed) |
the exact phrase that justifies the tag | customer's words
worth reusing in copy (or blank).
If a review fits no theme, tag it "other". Don't paraphrase
the justification; quote it.

Asking for the justifying phrase lets you spot-check twenty rows in a few minutes. If more than two or three are wrong, fix the frame's definitions and rerun, rather than correcting rows one by one.

#18What goes wrong, and how do you fix it?

#19How do you build a quote bank?

The most reusable output of review mining is a bank of customer phrases sorted by the job they do: describing the problem, describing the outcome, overcoming an objection. Ask Claude to pull them into a table with the source and date, then keep the file in your messaging project. Writers can then reach for real customer language instead of inventing it, and anything quoted publicly can be traced back and cleared.

#20Common questions

Can Claude analyse customer reviews? Yes. Upload reviews as a CSV, agree a coding frame of themes, and ask Claude to tag every comment and pull exact quotes for each theme.

Is it safe to upload customer data to Claude? Remove names, emails and other identifiers first, and check your plan's data terms and your company policy before uploading.

How many reviews can Claude handle? With code execution, Claude can analyse large CSV files by writing and running code. Spot-check the results against a sample either way.