AI-Powered Microcopy: Small Prompts, Big Conversions at Cartly

AI · 4 min read

Cartly, an e-commerce checkout provider, experimented with generative microcopy to reduce cart abandonment. Instead of generic CTAs and error messages, they used a constrained LLM pipeline to generate context-aware copy: emergency shipping reminders, payment failure explanations tailored to the user’s past behavior, and soft nudges when carts hit a threshold. The goal was clarity and relevance, not creative flair.

The team invested heavily in prompt engineering and safety constraints. Prompts included user attributes (region, last purchase), transaction context (cart size, shipping urgency), and strict style rules to keep copy concise and non-manipulative. Output passed through a rules engine to prevent illicit claims (e.g., guaranteed delivery times) and was reviewed by humans during a staged rollout to capture edge cases.

The result was a 6.5% relative uplift in completed checkouts where personalized microcopy was displayed, with no increase in support tickets. Lessons included the importance of narrow prompts, A/B testing for message tone, and an ongoing human-in-the-loop pipeline for monitoring drift and compliance as models evolved.