System Prompt
A system prompt is the instruction text an operator places ahead of a conversation with a language model, fixing the role, rules and output format the model follows for every turn that follows. The person typing questions does not see it, and their questions are read in its light.
What it fixes before the first question
The system prompt is where an operator says who the model is meant to be and what it may not do: the persona, the tone, the topics it declines, the language it answers in, the shape of its output. Anthropic’s prompting documentation describes giving the model a role this way, and the same slot can carry the rules for using tools: when to search the web, how many sources to read, how to cite them. A user message is one turn. The system prompt applies to all of them.
It is also where the model is told what to do when it is unsure. Whether an assistant hedges, names three businesses or names one is partly a property of the model and partly a rule written here by the operator.
Why it matters to a business that wants to be named
You do not write the system prompt of ChatGPT, Claude or Gemini and you cannot see it. What you can influence is the material the assistant retrieves under those rules. If the prompt says to search and cite, the retrieved pages are what the answer is built from, and a page that states a fact about a business in one place is what the citation rule can act on. Measuring visibility therefore means asking the assistant as a customer would, with its own system prompt in force, and recording what comes back; the methodology page describes that panel.
A common question
If a business cannot see or write the system prompt, what can it actually affect?
The material the assistant retrieves while working under those instructions. Nobody outside the operator writes or reads the prompt behind ChatGPT, Claude or Gemini, but when it says to search and to cite, the answer is built from whatever pages came back. A page stating a fact about a business in one place is what a citation rule can act on. That is also why visibility is measured by asking as a customer would, with the assistant's own prompt in force, and recording what returns.
Related terms
Sources
Role and instruction behavior: Anthropic’s documentation on system prompts. No figures on this page.