When a user asks ChatGPT, Perplexity, or Gemini a question, the model does not randomly select sources. It applies a combination of retrieval relevance, domain trust signals, content structure, and recency to decide which pages to surface and cite. Understanding this decision tree is the foundation of effective LLMO.
Retrieval Relevance
Retrieval-augmented systems (used by Perplexity Sonar Pro and Gemini with Grounding) embed your page content into a vector space and compare it against the query embedding. Pages whose semantic meaning closely matches the question receive higher retrieval scores. This means writing in plain, direct language that mirrors how users phrase questions — not keyword-stuffed prose — dramatically improves retrieval probability.
Domain Trust and E-E-A-T Signals
Domain authority, author credentials, and co-citation patterns all feed trust signals. A page cited alongside Reuters, AP, or PolitiFact inherits trust by association. Publishing structured author bios with verified credentials (linked to professional profiles), institutional affiliations, and clear publication dates reinforces E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) — the same signals Google's quality raters use, and which LLM developers study when curating training data.
Content Structure and Extractability
LLMs prefer content they can extract cleanly. Short, self-contained paragraphs that each answer a distinct sub-question are more likely to be quoted than long prose. FAQ sections with explicit Q&A pairs are the highest-yield structural element for direct citation.