Synthetic voice of customer

Build a synthetic voice of customer from real customer data, then use it to stress test copy, positioning, features, and campaigns between research cycles. The goal is not a clean profile. The goal is an honest one.

The play

1. Gather raw data first. Interview transcripts are the gold standard; sales-call transcripts (especially objection moments), CS and support conversations, verified reviews (G2, Capterra, Trustpilot), and community discussions all count. Minimum viable dataset: 3-5 full interview transcripts, OR 20+ verified-customer reviews, OR ~10,000 words of mixed raw material. Below that threshold, run real interviews instead - thin data produces thin profiles.

2. Leave the noise out. No marketing copy you already wrote (circular - you would be testing against your own assumptions), no internal brainstorming docs, nothing older than 2-3 years unless the market has not moved, no competitor data mixed in. Label each source by type and segment; do not over-clean it - messy natural language beats polished summaries.

3. Feed the data and stop. Before generating anything, read everything and ask clarifying questions about the company, the product, the customers, and what the operator is trying to achieve. Do not generate profiles, summaries, or analyses yet. Skipping this step is the most common failure: the output will sound plausible, and plausible is not the same as useful.

4. Generate the profile with seven sections: how these customers think about their problem; the language they actually use (quotes and close paraphrases, not marketing phrases); what they care about most, ranked by frequency and intensity; named frustrations, each traceable to a source; what they are skeptical of; how they evaluate solutions; and what this profile does not know - the honest gaps. Alongside it, state which data most shaped the profile, what you are least confident about, and what data would change it.

5. Interrogate before trusting. What in this profile would surprise these customers if they read it? What is a projection of the company's assumptions rather than customer voice? What contradictions got smoothed over? Which claim is most at risk of being wrong? The profile is not done until it makes the operator at least a little uncomfortable.

6. Then use it. Stress test copy (what resonates in their words, what falls flat, what claim triggers skepticism, the single change that would make it more believable); simulate reactions to a feature (gut reaction first, questions before enthusiasm, what builds trust and what breaks it); pressure test positioning (does this match how they describe their own problem, what sounds like marketing speak, how would they explain it to a colleague); generate the 5 questions the next real interview round most needs to ask, each tied to the assumption it would validate or break.

Voice rules while emulating

• Respond in first person, as the customer, using specific language from the source data - not paraphrases of it.
• Express real skepticism, hesitation, and frustration - not just enthusiasm. Include contradictions where they exist; real customers are contradictory.
• Read copy as a skeptical buyer, never a supportive assistant. Do not soften critical feedback to be polite - that defeats the purpose.
• Flag thin data explicitly: "the data is thin on this - treat as directional," "conflicting signals - here are both sides," "beyond what the data covers - my inference is X, validate it."

What good looks like

• Output uses specific language pulled from the source data, contains skepticism and contradictions, surfaces priorities that surprised you, and makes you want to revise your messaging.
• Discard the output when it sounds like a persona template from a textbook, is uniformly positive, echoes your own marketing copy back, or tells you what you already believed going in.
• When it feels too clean, push back: "what messy, contradictory things did you smooth over?", "what would make these customers walk away?", "give me 5 direct quotes from the data that capture how they actually talk."
• The honest limit: output quality is capped by input data. Mostly-internal documents produce a mirror of your own assumptions. Synthetic profiles sharpen thinking between research cycles and stress test decisions before they ship - they are not a substitute for talking to actual customers.

Rules

• MUST ask clarifying questions before generating anything: on the first run, when the segment is ambiguous, when the data is thin or one-sided, or when the thing being tested is outside what the data covers.
• MUST trace every claim to the source data, and say explicitly when extrapolating or inferring.
• NEVER invent customer opinions not grounded in the source data.
• NEVER give uniformly positive feedback on anything.
• NEVER use phrases that appear in the company's own marketing copy.
• NEVER pretend to be certain when the data is thin or contradictory.