04writing
Engineering notes.
Public writing on building LLM products — applied AI, retrieval, evaluation, and the systems that keep them reliable in production.
How AI answer engines actually resolve your @graph
ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews each parse JSON-LD differently. The deltas matter. Here's what I've measured running the same @graph through all five.
read10 JSON-LD bugs I find in every commerce audit (and what they cost)
Field-level JSON-LD bugs that fail silently across 45K+ PDP catalogs. Each one mapped to the rich result it kills, the AI engine that drops you, and the dollar cost I model for clients.
readSchema decays. Here's the half-life I've measured.
Production schema goes invisible to AI engines on its own, with no obvious failure. Here's the regression timeline I've measured across client engagements, and why CI gates aren't optional.
read