Claim
Every major AI search platform now routes queries through five sequential gatekeepers: Planner, Retriever, Reader, Critic, and Synthesizer. The Critic stage drops most content with no observable signal, making single-shot retrieval optimization obsolete.
Mechanism
Agentic RAG decomposes each query before touching content. The Planner breaks the query into sub-intents. The Retriever fetches candidates. The Reader parses what it fetches. The Critic evaluates fitness and drops content that fails its rubric, with no signal to the publisher. The Synthesizer assembles what survives. Content that passes the Critic does so because it matches the platform's internal rubric for relevance, authority, and format fit. Content that fails disappears silently.
Conditions
Holds when: a query is processed by any of the five named platforms (ChatGPT, Perplexity, Google AI Mode, Gemini Deep Research, Copilot Researcher). Each runs agentic RAG as of May 2026.
Fails when: a platform uses a simpler retrieval model without a Critic stage. Some enterprise search deployments and smaller AI tools may not run the full five-gatekeeper sequence.
Evidence
King maps the architecture across all five platforms and names the Critic as the invisible failure point. His diagnostic: run your content through each platform and observe where it drops out of the final answer.
"You cannot optimize what you cannot observe. Distilling your own version of it is the only path to durable GEO performance."
Signals
- Content appears in Retriever-stage tests (Perplexity source panel) but not in the final answer: Critic failure.
- Content passes on one platform and disappears on another: platform-specific Critic rubric gap.
- Shorter, structured content survives more consistently than long-form: Reader and Critic both favor parseable format.
Counter-evidence
King's model is based on observed behavior, not published platform specs. The internal architecture of each gatekeeper is inferred, not documented. Google AI Mode's Critic rubric may differ significantly from ChatGPT's. The framework is directionally correct but operationally approximate.
Cross-references
- Package the AEO audit as a year-long keep/start/stop deck (Mike King, AEO roadmap framing)
- Serve Markdown to AI crawlers you want indexing your content; Cloudflare measured an 80% token size reduction versus full HTML (format layer for LLM crawlers)