Claim
The best conversion copy is not written. It is assembled from customer language found in reviews, forums, and support transcripts. The LLM's job is pattern recognition and assembly. The human checkpoint is editorial: does the assembled language sound like the customer or like an AI approximation of one?
Mechanism
Customers writing about their own problems in forums and reviews use the exact words they would respond to in marketing. An LLM can recognize patterns in that raw language and structure them into landing page copy. The editorial review functions as a distinctiveness gate: anything that sounds like AI paraphrase fails the test before it reaches a buyer.
Conditions
Holds when: sufficient VOC data exists (reviews, forum threads, support transcripts, interview notes).
Fails when: the product is so new that customers have not yet generated written accounts of their own experience. Low-review-volume products lack the density to train the pattern.
Evidence
Cole:
"The world's best conversion copywriters don't write copy. They steal it."
The AI-native version treats customer language as the primary source and the LLM as the assembly layer. The human review determines whether the output passes or fails the authenticity test.
Signals
- Landing page phrases appear verbatim in customer review data
- Editorial review catches and removes AI-approximated language before deploy
- Copy assembled from VOC outperforms copy written from internal product briefs
Counter-evidence
AI assembly can produce fluent but generic output if the source VOC is sparse. Some markets have no public review data. In regulated industries, customer language may be too informal for the target audience.
Cross-references
- ins_voc-first-then-positioning
- ins_jtbd-interviews-surface-customer-language