Most SEO teams are still treating AI search like it's one channel. It isn't. ChatGPT, Gemini, and Perplexity have fundamentally different retrieval architectures, different training priors, and different content ingestion pipelines. That means the same website can be invisible to one engine and perfectly readable to another — not because of a ranking problem, but because of a structural one.
This isn't theoretical. We ran a live test across all three engines using a client site with a fully deployed JSON-LD entity graph. The results were a clean demonstration of exactly how differently these systems process the same URL.
The Test: Same Site, Three Different Answers
The site had a complete `@graph` schema: Organization, Person, LocalBusiness, `sameAs` arrays, `areaServed` nodes, and relationship links between a parent brand and regional subdomains. The structured data was clean. The entity graph was accurate.
We asked each engine the same question: describe this brand, its locations, and its relationships.
ChatGPT initially hallucinated a phone number that didn't exist on the site. When forced to re-read live text snippets, it corrected itself — but it still couldn't surface the `sameAs` schema arrays or the parent-child brand relationships defined in the JSON-LD.
Gemini mapped the entire enterprise relationship correctly on the first pass. It accurately parsed the brand graph, named the regional hubs, and even reproduced the JSON-LD structure when asked to output schema facts.
Perplexity refused to guess. It explicitly stated it couldn't verify raw JSON-LD schema nodes from its snippet payload, fell back to visible text items only, and labeled the structured data relationships as "unverifiable."
Three engines. Same site. Three completely different outputs.
Why This Happens: The Retrieval Architecture Gap
This isn't a bug. It's a feature of how each system was built.
ChatGPT: Prose-First, Schema-Blind
OpenAI's web search ingestion pipeline strips or heavily compresses `