For most of a 17-month U.S. clickstream study, 65%-85% of observed ChatGPT prompts did not match any term in a 27-billion-keyword search database. That does not mean these prompts lack intent; it exposes a measurement gap when conversational exploration, synthesis, and refinement are forced into taxonomies designed for short search queries.
Audience takeaways:
1. Separate underlying need, prompt expression, model synthesis, verification, and action in an Ask -> Refine -> Verify -> Act framework.
2. Audit when classic informational, navigational, commercial, and transactional labels remain useful - and when they create false precision.
3. Combine clickstream, surveys, diaries, and passive measurement to recover consumer-journey signals that become less visible inside AI interfaces.