AI startups face “reset risk”: each model breakthrough can instantly obsolete existing players. Standard SaaS metrics (CAC/LTV, retention) don’t capture this. The piece proposes a 4-layer framework to assess real defensibility: 1. Tech foundation – real performance delta, improvement velocity, and data edge. 2. Product depth – how deeply it solves the user’s job, embeds in workflow, and operates autonomously. 3. Market positioning – timing of capability thresholds, vertical vs horizontal strategy, and market structure. 4. Sustainability tests – what happens if GPT-N drops tomorrow, how costly it is to switch, and whether new users still come.