Graphite asks converting users which prompt they typed into which LLM

Ethan Smith, writing on LinkedIn on 15 October 2026, published a method Graphite now uses to collect real user prompts at conversion. The funnel asks three questions in sequence: how the user heard about them, which LLM they used, and the prompt they entered.

Smith reports that 30% of inbound leads at Graphite now self-report AI as the channel. Of those, 52% name ChatGPT and 40% name Claude. He also notes that AI-sourced leads tend to be less qualified than referral traffic.

Why last-touch misses most of it

Smith reports that ChatGPT sends 10x more referral traffic to Graphite than Claude, yet the self-reported lead share for Claude is almost as high. He reads this as Claude conversions being badly underreported by last-touch attribution.

The gap is wider across the whole AI channel. Only 21% of leads who self-report AI as the source show up as AI referral traffic in analytics. The other 79% arrived through search (49%) or direct (30%).

Malte Landwehr described the same two-step pattern on the podcast, suggesting an AI Search option in the standard "how did you hear about us" question, with a follow-up for the prompt itself (listen).

What to do with it

Before acting on anyone's prompt data, including your own, verify it on your site. Add the three questions to a conversion form and run it long enough to see whether your referral numbers understate AI the way Graphite's did. The gap is the finding worth checking, not the headline percentage.

Treat the prompts you collect as one input among several. They tell you which questions real buyers are asking an LLM when they end up converting with you. That informs which questions you should be the best answer to. It does not produce a list of pages to write against one by one, and the writing itself is still for the reader.