What is a system prompt?

99Visibility Research · Updated October 21, 2026 · 4 min read

Short answer

A system prompt is the set of hidden instructions an app gives an AI model before the user types anything. The company building the product (the integrator) writes it. It sets tone, rules, tools and limits. It can also steer whether the assistant searches the web, so the same underlying model can answer the same question differently in two products.

Key takeaways

A system prompt is written by the app builder, not the user, and the user usually never sees it.
Integrators can use it to steer when an assistant searches the web, which changes whether your site can be retrieved at all.
The same model can behave differently in a consumer app, an API test and a third-party product because each runs different instructions and settings.
Publishers cannot edit or override another app's system prompt, and hidden text that tries to is illegitimate and risky.
Any AI visibility measurement should state which surface and retrieval state it used, because the system prompt is part of the setup.

Definition

A system prompt (also called system instructions or the integrator prompt) is text placed in front of every conversation by whoever deploys the model. It tells the model who it is, what it may do, which tools it has, and how to format answers.

Three parties shape an AI answer:

The model vendor trains the model and sets its defaults.
The integrator builds the app and writes the system prompt.
The user types the question.

The user sees only their own message. The system prompt runs underneath it. For a full walk-through of why one model gives different answers in different products, see why Claude answers differently in different apps.

How it works

When an app sends a request to a model, it bundles the system prompt with the user's message and any tool definitions. The model reads all of it together and decides what to do.

Web search is one of those tools, and the decision to use it is not fixed. As of October 2026, the evidence shows:

Claude: the model decides when to search, and the integrator's system prompt steers that decision. In Claude.ai, search is a user toggle with no published default.
ChatGPT: search is model-discretionary. OpenAI says ChatGPT typically rewrites your query into one or more targeted queries, and may use saved memories when doing so.
Perplexity: on its Agent API, web search must be opted in. Code that never turns it on gets answers with no web grounding.

So an integrator can write instructions that make search more or less likely, or remove the search tool entirely. The model is the same. The behavior is not.

Why it matters for AI visibility

Your pages can only be cited if the assistant retrieves them. If the system prompt discourages search, or the app never offers the search tool, the model answers from what it learned in training. Nothing you publish this quarter reaches that answer.

This has three practical consequences.

Results don't transfer between products. A model that names your brand in one app may not in another, because the instructions and tools differ.
API tests aren't app tests. A test that calls a model through its API with search off measures a different setup from a consumer app that searches at its own discretion. Ask any tool which setup it used.
You can't see the real setup. Outside parties don't see an app's system prompt or its rewritten queries. That is why AI visibility tools can't see the prompts people actually type.

The next move: control what you can. Make sure retrieval bots can reach your pages and that key facts are in the HTML. Then measure each surface separately and record whether search was on. A reproducible ChatGPT tracking method shows how to hold those variables steady.

Common misconceptions

"I can optimize for the system prompt"

You can't. You don't know what it says, and it changes when the integrator updates it. No published study shows a page technique that reliably survives across apps.

"Hidden text can override it"

Some pages hide instructions aimed at AI models, hoping to override the app's rules. This is prompt injection, and it is illegitimate. Research on LLM search agents (arXiv:2606.16821, June 2026) found that the same deployment scaffold could raise or lower attack success depending on the model behind it. That makes the effect unpredictable, and the reputational risk is real. See hidden text for why it backfires.

"The system prompt is the same as the user prompt"

It isn't. The user prompt is the question. The system prompt is the standing instruction set that frames every question.

"One test tells me how 'the model' sees my brand"

It tells you how one product, with one set of instructions and one retrieval state, answered on one day.

Frequently asked questions

Can I see an app's system prompt?

Usually not. Integrators rarely publish their system prompts, and they can change them at any time. Some details leak out through vendor documentation, such as whether search is a user toggle in a given app. Treat any system prompt you find online as a snapshot of one product on one date, not a stable rule you can build a strategy around.

Does the system prompt decide whether my site gets cited?

It influences whether search happens at all, which comes first. If no search runs, nothing from the live web can be cited. If search does run, retrieval and the model's own choices decide which pages get used. You control neither the system prompt nor those choices. You do control whether retrieval bots can reach your pages and read your key facts.

Why does ChatGPT answer differently in the app than through the API?

The app and the API can run different system instructions, different default models and different search behavior. OpenAI says ChatGPT search typically rewrites your question into targeted queries and may use saved memories. An API call with search off skips all of that. Compare like with like, and record the model, search setting and date for every test.

Can a competitor change what an AI says by editing a system prompt?

Not another company's system prompt. A competitor could publish pages with injected instructions, hoping an assistant retrieves them. That is manipulation, and its success varies by model. If you suspect it, read whether a [competitor can manipulate AI answers](/questions/competitor-manipulate-ai-answers) and scan third-party pages that mention your brand for injected text.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.