What Is a System Prompt?

A system prompt is the standing instruction set an AI provider injects before every user conversation, defining the assistant's persona, rules, tool-use behavior, and answer formatting. Users never write it and rarely see it, yet it shapes every response — including whether the model searches the web, how many sources it reads, and how it credits them.

What system prompts actually control

Beyond tone and safety rules, production system prompts govern the mechanics GEO practitioners care about: search-trigger criteria (when a question warrants live retrieval versus parametric recall), source-handling instructions (prefer recent results, verify claims across multiple sources, avoid low-quality domains), citation formatting (inline markers, source panels, link styles), and refusal policies around sensitive or commercial topics. In effect, the system prompt is the editorial policy of an answer engine, written in natural language.

What we know from public examples

Anthropic openly publishes Claude's system prompts alongside model releases — an unusual transparency practice that confirms how detailed these instructions get, down to guidance on hedging, formatting, and when to search. For other engines, knowledge comes from prompt-extraction attempts and observed behavior shifts. What is consistent across engines: search behavior and citation style are instructed, not emergent, which explains why an engine's sourcing habits can change overnight without any model swap.

Why this matters for visibility work

Three practical consequences. First, engine behavior is a moving target — a revised system prompt can alter citation frequency, answer length, or brand-naming caution while the underlying model stays identical, so attribute visibility shifts carefully. Second, engines differ by policy, not just capability: one engine's instructions may favor heavy citation while another's favor synthesis with minimal linking, which is why per-engine strategies outperform uniform ones — see the ChatGPT visibility guide for one engine's specifics. Third, you cannot optimize the system prompt itself; you can only make your content the easiest thing for its instructions to reward: current, verifiable, clearly attributed. The glossary entry on retrieval bias covers the source-selection tendencies these prompts encode.