In a previous article on this site, I ran a live test that exposed the exact architectural divide between how ChatGPT, Claude, and Gemini handle JSON-LD schema markup. The short version: Gemini reads your schema through Google's index layer. ChatGPT strips script blocks from its context window entirely. Claude won't engage with URLs that lack an established index footprint.
That test answered the diagnostic question. This article answers the implementation question: given that different AI retrieval systems operate at different layers, what do you actually do about it?
The answer is the dual-channel approach. You keep your JSON-LD schema exactly as it is — it's doing real work for Gemini, Google AI Mode, and Bing Copilot. And you add a prose layer that mirrors the same critical information in visible text, so ChatGPT, Claude, and Perplexity can extract it through direct fetch.
This isn't redundancy. It's two channels serving two different retrieval architectures. Here's the complete checklist.
Why Two Channels Are Required
Before getting into the implementation, it's worth being precise about why this matters.
The OtterlyAI GEO experiment, which ran 319 prompts across seven AI platforms over three months, found that six out of seven platforms could not fetch or correctly interpret schema markup when directly asked. Only Gemini succeeded. Google AI Mode partially succeeded but hallucinated schema types that weren't on the page. ChatGPT, Claude, Perplexity, and Microsoft Copilot all failed to access the raw JSON-LD.
The reason is architectural. Most AI retrieval pipelines use HTML-to-Markdown extraction workflows that either ignore or strip `