| url | https://www.nber.org/papers/w34255 |
|---|
The first rigorous, large-scale study of consumer ChatGPT usage, based on internal OpenAI data from May 2024 through June 2025. Notable for actually answering a question that was, until this paper, almost entirely guesswork: what are people doing with this thing?
The headline finding
The killer use case for ChatGPT isn’t what most observers (or OpenAI) initially thought. The breakdown skews dramatically toward non-work, practical guidance, writing assistance, and information seeking — not enterprise productivity, not coding, not “agents.” Over 70% of consumer messages are non-work-related.
Translated: ChatGPT became the fastest-growing consumer app in history because people wanted a smart friend to think out loud with. Everything else is a rounding error on top of that core use case.
Why it matters here
This paper is the empirical answer to the question Vohra would ask of ChatGPT under his PMF engine: who are the very disappointed users, and why do they love the product? The answer turns out to be “almost everyone, for everyday cognitive offload.”
That has big implications for the ChatGPT PMF case study: the atomic unit that worked was the conversation, not the agent or the plugin. And it explains why every attempt to “move ChatGPT up the value chain” toward agents or workflows has been slower than expected — the existing product is already solving the job most users hired it for.
See also product-market-fit and atomic-concepts.