LLM Technical PhenomenaEvidence Map for GEO/AEO Strategy

Research-backed intelligence for optimizing content in the age of AI search

🎲

Non-Determinism

Understanding why LLMs produce different outputs for the same prompt - KV cache batching, MoE routing, and mitigation strategies.

📉

Context Rot

Why long-context performance degrades 20-60% and how the "lost in the middle" phenomenon affects your content.

⚖️

Alignment Tax

How RLHF training suppresses 5-15% of valid content and which categories are most affected.

🎯

Position Effects

Why content position matters 2-3x more than you think and how to front-load for maximum citation probability.

🔒

Security Considerations

Prompt injection vulnerabilities and the MCP standard for safer multi-turn conversations.

🔮

Interpretability Gaps

Why we can't fully explain LLM behavior and what that means for future-proofing your strategy.

Fresh Updated January 2026

About This Documentation

This documentation synthesizes 2025 research from Thinking Machines Lab, Chroma Research, and ACL into actionable Standard Operating Procedures (SOPs) for SEO professionals transitioning to:

GEO (Generative Engine Optimization)
AEO (AI Engine Optimization)

Each phenomenon is grounded in peer-reviewed research, empirical data, and field validation.

Quick Stats

Metric	Value	Source
Platform citation overlap	42%	Cross-platform analysis
Context rot degradation	20-60%	Chroma 2025 (18 models)
Middle content penalty	30% worse	Lost in the Middle study
Deterministic inference cost	2x slowdown	Thinking Machines Lab
RLHF performance drop	5-15%	ACL 2024

Research Sources

Thinking Machines Lab (2025): "Defeating Nondeterminism in LLM Inference"
Chroma (2025): "Context Rot: How Increasing Input Tokens Impacts LLM Performance"
152334h (2024): "Non-determinism in GPT-4 is caused by Sparse MoE"
ACL 2024: "Mitigating the Alignment Tax of RLHF"

Original Source: hiddendriftstate.com / Novel Cognition 2025