System Prompt

Before a user types anything, the platform loads a system prompt into the context window: identity, formatting rules, safety boundaries, and operational instructions, when to invoke web search, how many sources to cite, how to handle recommendations and comparisons. The model treats these instructions as higher priority than user messages.

System prompts are product decisions, updated frequently and rarely published. Two platforms running the same base model can answer very differently purely because of this layer.

Instructions like "present balanced options," "cite sources for factual claims," or "avoid promotional language" directly govern brand treatment: whether you get a flat list of five competitors or an opinionated single pick, whether claims carry citations, and how hedged recommendations sound. Search-trigger rules decide if an answer uses live retrieval, where your current content can win, or frozen training memory.

When a platform revises its system prompt, answer formats and mention patterns can shift overnight across all users, with no model change at all.

You cannot read or edit platform system prompts, but you can align with their consistent tendencies: models are broadly instructed toward sourced, neutral, comparative answers, so citable evidence, third-party validation, and honest comparison content fit the format engines are told to produce. Promotional copy fights the instructions; verifiable claims work with them.

Sudden platform-wide shifts in answer style or mention patterns often signal a system-prompt update rather than anything you did. Continuous tracking through Geonimo helps separate those platform-side changes from genuine movements in your own AI visibility.