Every major platform transition has produced a metrics crisis. Web 1.0 had 'hits'. Social had MAU and DAU. Smartphones confused unit sales with installed base. Now generative AI has the same problem, and the people reporting numbers know it. OpenAI publishes 'weekly active users' occasionally, in round numbers. Sam Altman co-founded a social company. He knows WAU is a weak signal. If someone uses a product once a week, it is not load-bearing in their life.
The definitional rot goes deeper than cadence. Survey data asking whether consumers or enterprises 'use AI' is nearly worthless. The question conflates ChatGPT with Alexa and Snapchat filters, all technically machine learning, all historically called AI. On the enterprise side, the gap between 'someone in marketing uses Midjourney sometimes' and 'we rebuilt invoice processing on an LLM' is the difference between a rounding error and a structural bet. Google and Microsoft have pivoted to reporting tokens generated, which is the 1996 equivalent of reporting backbone bandwidth: directionally real, analytically hollow. Model efficiency gains, agentic workflows, media generation, and Google force-feeding AI Overviews to every search user all move that number in different directions simultaneously.
The original piece works through each metric layer by layer, using the FTC's Instagram-versus-TikTok framing as a live case study in how metric choice determines competitive reality. The token-generation chart problem alone is worth reading in full: the author maps it onto YouTube bandwidth circa 2005, where you would need to decompose users, views per user, video length, completion rate, and resolution before the number meant anything. That decomposition does not yet exist for AI. That is the actual story.
[READ ORIGINAL →]