A system prompt is an instruction block placed at the start of a model's context by the application, not by the user. It tells the model who it is acting as, what it may and may not do, which tools it can call and how to format answers. In AI search products, the system prompt and related instructions influence things like how many sources are cited, how recommendations are presented and how cautious the answer is.
What system prompts typically contain
Providers rarely publish their full production system prompts, and they change often. Developer documentation from model vendors describes the general pattern, which usually covers:
- Role and tone: for example, a helpful assistant that writes concisely
- Rules: topics to avoid, safety policies, formatting requirements
- Tool definitions: search, browsing, code execution, shopping
- Citation behavior: when and how to attribute sources
- Context: current date, user location or preferences, where allowed
Why it matters for brand visibility
The same underlying model can produce different answers inside different products, partly because each product wraps it in different instructions and tools. This is one reason a brand can be recommended in one assistant and missing from another that uses a related model. It is also why reverse-engineering visibility from one product's behavior does not always transfer.
For your own AI features, such as a support bot or in-app assistant, the system prompt is where you set accuracy rules: answer only from provided documents, say when you do not know, and link to sources.
Example
A support bot that oversold
Driftmoor, a fictional logistics SaaS, launches a support assistant. Early transcripts show it promising features on the roadmap as if they were live. Adding system prompt rules to answer only from the current help center, and to say 'not currently supported' when nothing matches, fixes most cases.
Common mistakes
Assuming you can influence another product's system prompt
You cannot. Influence comes from the content the model retrieves and the information it learned in training.
Putting secrets in your own system prompts
System prompts can sometimes be extracted by users. Do not include keys, internal data or anything you would not want exposed.
Questions
Normally not, but users sometimes coax models into revealing them. Treat a system prompt as potentially visible.
Custom instructions set by a user are layered into the context in a similar way, but the provider's system prompt sits above them and usually takes priority.
