How Do I Build a Customer-Voice Swipe File From Reviews in an Afternoon?

Most ecommerce marketers I talk to write ad copy from memory. They know the product, they know the brand, and they write from that internal frame until the copy sounds polished and sounds nothing like how customers actually describe the problem they needed solved. The ads get clicks from people who already understand the product. The people who need it most, the ones who are typing their pain into a search bar at 11pm, scroll right past.

The fix is a swipe file built from real customer language. Not from your brand guide, not from competitor taglines, but from the actual words customers chose when they were not trying to be marketing-friendly. That language lives in public reviews, and AI can pull it out, categorize it, and turn it into usable copy anchors in a single afternoon session.

This guide walks through the whole workflow: sourcing reviews ethically from public data, running the extraction prompt, separating pain from desire from objections, pulling the verbatim phrases worth keeping, and turning those anchors into testable headlines. By the end, you will have a structured document your copywriter or agency can use as the research layer for every campaign you run this quarter.

Why this matters for ecommerce and DTC brands specifically

Ecommerce copy lives and dies on specificity. A skincare brand that says "feel confident in your skin again" is competing on the same line as a thousand other brands. A skincare brand that opens with "I used to plan my whole morning around hiding my forehead" is speaking to the person who wrote that review at 2am after another frustrating morning. The second line wins because it sounds like something a customer actually thought, not something a marketer wrote.

The problem: sourcing that language manually takes hours. Reading through 200 reviews on Amazon or Trustpilot, highlighting the phrases that resonate, sorting them into themes, writing them up into a usable format is a day of work a marketing manager does not have. Most brands do it once, maybe twice a year, and the output gets used for one campaign launch before going stale. AI collapses that work to 25 minutes and makes it repeatable on every product line and every new campaign cycle.

What an AI extraction workflow actually does

An AI extraction workflow treats the chat interface as an analyst. You feed it a corpus of raw text (in this case, review content), give it specific instructions on what to find and how to categorize it, and get back structured output you can work with directly.

Three things make this different from just reading reviews yourself:

• Scale. AI reads 50 reviews in seconds and surfaces the patterns across all of them. You read 50 reviews and remember the last five.
• Categorization. AI sorts pain, desire, and objections into separate buckets in one pass. Manual sorting is tedious enough that most marketers skip it.
• Verbatim fidelity. Told explicitly to pull phrases as written (not summarize them), AI returns the specific words reviewers used. That distinction matters.

Think of it as a research assistant who has read everything and taken notes in the format you asked for.

Before you start

You need:

• A free account on Claude (claude.ai) or ChatGPT. Either works for this workflow.
• A product with at least 20 public reviews on Amazon, Google, Trustpilot, G2, or a similar platform. If your own product has fewer reviews, competitor reviews from the same category work just as well.
• 25 to 30 minutes of uninterrupted time. The workflow is sequential and fast, but switching context mid-session costs you the thread.
• A Google Doc or Notion page open to paste the output into as you go.

Before you paste anything: review text from public pages is safe to use for research. Personally identifiable information (reviewer names, email addresses, account data behind a login) is a different matter. GDPR, CCPA, and FTC ad rules all have implications for what you do with customer data. We have a dedicated section on this below. It is non-negotiable.

Task 1: Source the corpus ethically from public reviews

The failure pattern here is trying to do too much. Marketers who hear "build a swipe file from reviews" often start thinking about scraping tools, automated pipelines, or pulling data at scale. For a 25-minute session, you do not need any of that. You need 30 to 50 reviews in a text document.

The fastest way to get them: open the product page on Amazon, Trustpilot, G2, or whichever public review platform your product or competitor's product uses. Read through the reviews, copy and paste the text of the ones that seem detailed (more than two sentences, specific about a problem or result). Do this for 10 minutes. Aim for 35 to 50 reviews. Do not worry about filtering them yet.

What to look for when selecting reviews to copy:

• Reviews that describe a situation before the product ("I had tried everything for six months before I found this")
• Reviews that describe a specific result with specific language ("my inbox went from 300 unread to zero in two days")
• Reviews that push back on something ("I was skeptical because every other app claims the same thing")
• Reviews that answer an objection ("I thought the price was too high until I did the math on what I was spending on returns")

Skip one-sentence reviews, reviews that just list star ratings with no text, and reviews that read as brand-planted. You want raw customer experience, not polished summaries.

One firm boundary: public-facing reviews visible without a login are fair sourcing for market research. Do not scrape data from behind authentication walls, do not pull from private communities or closed Facebook groups, and do not pull review data from competitor CRM or internal customer records you might have access to through a platform integration. Public only.

Paste everything into a single text block in your Google Doc. Label it "Raw reviews: [Product Name] [Date]." You are ready for the extraction step.

Task 2: Run the extraction prompt on the raw corpus

The extraction prompt is doing one thing: pulling the specific phrases customers used, categorized into three buckets. The common failure at this step is asking AI to summarize or paraphrase the reviews rather than extract verbatims. Summarized output gives you marketing language. Verbatim extraction gives you customer language. The difference is the whole point.

What to ask AI for instead:

You are a direct-response copywriter doing voice-of-customer research. Below are [X] customer reviews for [Product Name], a [brief product description] sold to [audience]. Read all of them.

Pull exact phrases and partial sentences from the reviews as written by the customers. Do not summarize or paraphrase. Pull the actual words.

Organize your output into three labeled sections:

PAIN: Phrases that describe the problem before the product. What was broken, frustrating, or failing. DESIRE: Phrases that describe the outcome or transformation the customer wanted or got. What they hoped for or achieved. OBJECTION: Phrases that express doubt, hesitation, or skepticism the customer had before buying or overcame after buying.

For each section, list 8 to 12 verbatim phrases. If a phrase appears across multiple reviews, note it as "repeated" and include it once.

[Paste your 35 to 50 reviews below this line]

The output from this prompt is the core of your swipe file. Review it once before moving on. The phrases that make you think "that's exactly how our customer talks" are the ones to flag. The phrases that feel generic or vague, even if technically pulled from reviews, are the ones to leave aside.

One useful variant for brands with adjacent audiences: run the same extraction on reviews from a product that solves the same problem in a different category. A project management tool that sells to marketing agencies can mine reviews from time-tracking software and enterprise email tools for the category-level pain language that crosses products.

Task 3: Separate pain from desire from objection

The three buckets are not just organizational tidiness. They map to different ad functions.

Pain language works in the first three seconds of an ad, the hook. It signals to the right person that this ad is about them. A pain-first hook filters out people who do not have the problem and pulls in the people who do.

Desire language works in the middle of the ad and in the headline of the landing page. It describes the world after the product, which is what the buyer is actually buying. People do not buy project management software; they buy the feeling of a team that is not constantly on fire.

Objection language works in the close, in FAQ sections, and in retargeting copy aimed at people who visited but did not buy. When a customer writes "I was skeptical about the monthly price but it paid for itself in the first week," that is the answer to the exact objection that is keeping unconverted visitors from buying.

If the extraction prompt output mixes the three, run this secondary sort:

Here is the verbatim extraction from the customer reviews. Review the phrases in each section and check my categorization. Move any phrase from PAIN to DESIRE or OBJECTION if it more accurately fits there. Add a label to each phrase: (Hook) if it would work as an ad opening, (Landing) if it describes an outcome suitable for a landing page headline, (Retargeting) if it addresses a hesitation or price objection.

[Paste the extraction output]

The labeled output is what you will hand to your copywriter as the structured swipe file. Three sections, phrases labeled by function, verbatims intact.

Task 4: Pull the verbatim voice anchors worth keeping

Not all verbatims are equal. Some phrases are technically from customers but express ideas in ways that are too personal, too niche, or too specific to one reviewer's situation to generalize. Others are gold because they compress a common feeling into words your creative team could never write on their own.

A voice anchor is a phrase that meets three tests: it is specific, it is emotionally direct, and it could belong to any of your target customers rather than just the one who wrote it.

Run this prompt on the sorted output to surface the top anchors:

From the phrases below, identify the 6 to 8 that have the highest potential as voice anchors for direct-response ad copy. A voice anchor is a phrase that is specific (not generic), emotionally direct (shows feeling, not just logic), and generalizable (could apply to most of the target audience, not just one person's niche situation). For each anchor you select, give one sentence on why it works.

[Paste the labeled extraction output]

The output from this step is the part of your swipe file you will actually use in headlines. Six to eight phrases, each flagged with a brief explanation of why it works. For a DTC supplement brand, an anchor might be "my knees stopped screaming at me by week two." For a B2B SaaS product, it might be "I stopped running the same status meeting four times a week." Both are specific, both are emotionally direct, and both apply to a wide slice of the target audience.

If your extraction produced fewer than six strong anchors, the corpus was too small or too shallow. Go back and add 15 to 20 more substantive reviews before running the anchor selection again.

Task 5: Turn voice anchors into testable headlines

The swipe file is the research. The headlines are the output. This step turns the anchors into actual copy your team can test in ads and on landing pages.

What to ask AI for instead of writing headlines from scratch:

You are a direct-response copywriter. Use the voice anchors below as the raw material for ad headlines and landing page headlines for [Product Name]. The target audience is [brief audience description].

For each anchor, write 3 headline variants:

• One that uses the anchor language almost verbatim as the hook
• One that reframes the anchor as a question
• One that leads with the outcome implied by the anchor

Headlines should be under 12 words each. Do not use generic marketing language. Keep the specificity and emotional directness of the original anchor.

Voice anchors: [paste your 6 to 8 selected anchors]

The output is 18 to 24 draft headlines. Most will need light editing. Some will be strong enough to test as-is. The ones that land closest to the original customer language, with minimal editorializing, are usually the ones that perform best.

For paid social specifically: the question-format variant often outperforms the statement format on scroll-heavy platforms. "Still planning your whole morning around hiding your forehead?" stops the scroll better than "Stop hiding your forehead" because it creates pattern recognition rather than issuing a command. Test both. Let the data tell you which your audience responds to.

The ecommerce-specific prompts that actually work

After running this workflow with DTC and ecommerce brands across a range of categories, four prompt moves separate useful output from generic output.

Specify the product and audience in the opening line. "You are analyzing reviews for a DTC knee supplement targeting active adults over 40" produces different output than "You are analyzing product reviews." The second prompt produces category-level language. The first produces audience-specific language. The difference shows up immediately in the pain phrases extracted.

Demand verbatims, not summaries. If you do not explicitly say "pull exact phrases as written, do not paraphrase," AI defaults to summarizing. Summaries give you the idea. Verbatims give you the words. Every extraction prompt in this guide includes an explicit instruction not to paraphrase. Keep it there.

Name the ad function for each category. Telling AI that pain phrases are for hooks, desire phrases are for landing pages, and objection phrases are for retargeting gives the output a functional shape rather than a research-paper shape. It also helps your copywriter know where to use each phrase without needing an explanation.

Constrain the anchor count. Asking for the top 6 to 8 anchors forces AI to rank rather than list. Unlimited output is a list of 40 phrases that all look equal. Constrained output forces a quality judgment. The constraint also makes the handoff to your creative team cleaner: six to eight phrases fits on a single reference card.

The GDPR, CCPA, and FTC ad rules non-negotiables

This section is short because the rules are clear, but it is the most important section in this guide.

Do not put any of the following into a consumer AI tool:

• Customer email addresses, names, or any personally identifiable information linked to review data
• Internal CRM data, purchase history, or customer records pulled from your own platform
• Review content from platforms that require a login to access (gated data)
• Review data purchased from a third-party data broker without verifying the terms of use
• Customer support transcripts, even anonymized, without explicit consent language in your terms
• Any data from EU-based customers outside of GDPR-compliant handling
• Any review text where the reviewer has explicitly requested removal or opted out

The workflow in this guide uses publicly visible review text only. Review text on public pages like Amazon, Google, and Trustpilot does not constitute personal data under GDPR or CCPA in the same way that a customer's email address does, because it was published publicly by the reviewer. That said, if you are building an automated pipeline (rather than the manual paste-and-run workflow in this guide), you need a lawyer to review the scraping terms of the platforms you use and the GDPR applicability to your specific data flows.

On the FTC side: using customer review language to inspire your copy is standard market research. Quoting a specific reviewer in an ad and attributing the quote to them is an endorsement and subject to FTC endorsement rules. The workflow in this guide produces copy anchors that your team writes from, not direct lifts attributed to named reviewers. Keep that distinction clear in how the swipe file gets used downstream.

If your brand has signed a Business or Enterprise AI agreement with a Data Processing Addendum, the rules on what you can input may be broader. Ask your privacy counsel or compliance lead what that agreement covers. Do not assume.

When NOT to use this workflow

This workflow is the right tool for a specific problem. It is not the right tool for everything.

• Do not use it to fabricate customer language. If reviews are thin or nonexistent, the output will reflect that. AI filling in the gaps based on what it thinks customers would say produces marketing-speak, not customer voice. Wait until you have real reviews to mine, or use competitor reviews with the sourcing labeled clearly.
• Do not use it as a substitute for actual customer conversations. Review mining gives you language for ad copy. It does not replace customer interviews, user testing, or post-purchase surveys. Reviews are already-filtered opinions from self-selecting customers. The people who did not buy never wrote a review. Qualitative research fills that gap.
• Do not paste review text that includes personal data. If a reviewer left their email, phone number, or full name in the review body (it happens), remove it before pasting into the prompt. The workflow requires only the substantive review content.
• Do not use this workflow to attribute quotes to specific reviewers in ads. Using a verbatim phrase from a review in an ad and attributing it to a named person ("Jane from Austin says...") without that person's consent triggers FTC endorsement rules. Use the language to inform your copy, not to create attributed testimonials.

A simple rule: this workflow is an unfair advantage on the 80% of copy research where you need to know how customers talk. Trust direct customer conversations and consent-based testimonial collection for the 20% where attribution and legal compliance have real weight.

The quick-start template

Copy this scaffold, fill in the brackets, paste into Claude or ChatGPT:

You are a direct-response copywriter doing voice-of-customer research for [Brand Name]. The product is [one-sentence product description]. The target audience is [audience descriptor: age range, situation, specific pain they are trying to solve].

Below are [X] public customer reviews. Pull exact phrases as written. Do not paraphrase.

Organize output into three sections: PAIN (problem before the product) DESIRE (outcome or transformation wanted or achieved) OBJECTION (doubt or hesitation expressed or overcome)

8 to 12 verbatims per section. Note any phrases that repeat across reviews.

After the extraction, identify the 6 best voice anchors: specific, emotionally direct, generalizable to most of the target audience. For each, write 2 headline variants under 12 words.

[Paste reviews here]

For recurring use: save this template in your team's creative brief doc or Notion template. Every new product launch or campaign refresh runs the same scaffold against the current review corpus. The swipe file updates in 25 minutes instead of a full research day.

Bigger wins beyond the basic swipe file

Once you have run the workflow once and seen what the output looks like, the next round of value comes from building it into regular creative practice.

A quarterly review pulse. Run the workflow every quarter on your top three SKUs and one key competitor. Customer language shifts as the category matures, as competitors change their messaging, and as new buyers enter the market. A quarterly swipe file refresh means your ad copy stays calibrated to how customers are talking now, not how they were talking 18 months ago at launch.

A category-level swipe file for the whole brand. Run the workflow across all of your products and three to five competitors at once. The output becomes a category-wide voice document that your creative agency or freelance copywriters can use as their starting brief for any campaign. Agencies that get this document produce better first drafts because they are starting from real customer language instead of brand-guideline abstractions.

A landing page rewrite anchored in voice. Take the top three pain verbatims and the top three desire verbatims from the swipe file and rewrite your product page around them. Swap the current headline for the top voice anchor. Rewrite the first paragraph in the language customers use to describe the problem. Replace generic benefit bullets with outcome phrases pulled directly from the desire section. This is the highest-value application of the swipe file because it improves conversion on traffic you are already sending.

For a deeper look at how AI fits into the broader content and marketing workflow, the companion reading is the Content Volume Paradox white paper. It covers why more AI content is not automatically better content, and how brands that use AI as a research and voice tool (rather than a volume machine) tend to outperform on both rankings and conversion.

The ecommerce AI consulting connection

Building a swipe file from reviews is one tactic inside a larger question: how does an ecommerce or DTC brand use AI across the marketing, creative, and operations stack without producing content that hurts performance by overwhelming search and social algorithms with undifferentiated volume? That question does not have a one-size answer. It depends on the brand's current stage, the category's competitive intensity, and the quality of the underlying product and customer data.

The AI Consulting for E-commerce page covers the full picture: where AI is consistently producing ROI for ecommerce brands right now, where it is producing noise that costs more than it saves, and what an engagement looks like when it is scoped correctly for a brand at your stage.

Closing

The point of this workflow is not faster copywriting. The point is that your ads stop sounding like ads and start sounding like something a real person would say to a friend who had the same problem. That shift is what drops cost-per-click and raises conversion rates at the same time, because the same language that attracts the right person also repels the wrong one.

Run the workflow on one product this afternoon. Pick the review source, run the extraction prompt, pull the voice anchors, write three headlines. Then put one of those headlines in your next ad set as a test against your current control. The data from that test tells you whether the workflow is worth making a quarterly habit.

If you want to talk about how AI fits into your ecommerce marketing operation at the program level, the AI Consulting for E-commerce page lays out the full picture and how an engagement works.