What AI Actually Sees: System Prompts, Context, and AI Visibility Tracking
Your prompt is only one part of what an LLM actually reads. Here's what really shapes AI answers about your brand, and why most visibility tracking measures an output no buyer ever sees.
When you submit a prompt to ChatGPT, Claude, Gemini - whatever you have adopted as your model of choice - your prompt is not the only thing that is actually given to the model. This article explains what happens under the hood when you pose a question to an LLM, how it influences the response you get, and why that is so important for AEO professionals and marketers to understand.
One useful distinction I would like to establish at the outset is the difference between the model and the product, as I will refer to both intentionally. The model is the underlying intelligence. Think of it like the backend or database. The product is the interface through which you engage with that model, which typically carries additional features to help the user extract quality, context-aware outputs.
What influences LLM responses
When you start a conversation with an LLM, your prompt is only one part of the actual query that is sent to the model. The product does extra work to help you get a good output from the model.
The user prompt is combined with a system prompt, as well as various layers of context, to make up the overall input into the model. That input is comprised of:
User prompt. Your specific input.
System prompt. A fixed instruction block written by the provider (OpenAI, Anthropic, Google) and prepended to every single conversation. The user never sees it, but it establishes the model’s identity and voice, its knowledge cutoff, behavioural patterns, formatting rules, some of its safety guidelines and the tools at its disposal to answer (e.g. connectors, search). It is identical for everyone.
Context. The additional user-specific information and instructions passed alongside the prompt.
- Configuration. Explicit standing instructions you have set in the product such as behavioural preferences and personal or professional context.
- Memory. Implicit preferences and facts the product has picked up about you during conversations.
- Conversation History. Preceding messages in the current thread.
- Metadata. Contextual data such as location, language, time.
- Documents. External content either attached by the user or fetched by retrieval. We wrote a separate explainer on retrieval, and the role it plays, here.
Why this matters for AEO, and why it should impact the visibility tracking vendor you choose.
This is an element of AI search tracking that is hugely overlooked, and likely causing many brands to carry forward an extremely misguided view of their AI visibility.
The majority of AEO tools’ visibility tracking remains extremely straightforward. Run a set list of prompts, pick out brand mentions and citations in the responses, and surface that in a dashboard. This approach falls down when you realise that two users can run the exact same prompt and get completely different outputs.
Tools generally operate in one of two ways.
- Call the API directly. This omits the system prompt and the context - everything AI products build to enhance the outputs your buyers actually receive. It is the most basic, most vanilla, least tailored output these models are capable of producing.
- Run prompts in the browser via a virtual CPU. This is certainly more sophisticated, and in some cases, an excellent idea to simulate the system prompt. The issue is that the system prompt is not the influential bit - it governs personality, behaviour and voice primarily. What matters to the substance of AI responses is context, and in that sense, these tools contextualise prompts for the lowest common denominator. They simulate the average consumer experience, to make their outputs broadly applicable. But your B2B buyers don’t get the average consumer experience, they get context-aware outputs that reflect their unique circumstances.
If you’re Colgate or Nike, it’s a good thing - you want your tracking to reflect the most common consumer experience. If you’re a specialist B2B brand, it’s a big problem, because your entire value proposition is built around being a specialist tool for a specific type of buyer.
As your buyers become more and more proficient with AI and understanding how to tune and configure AI-based products to achieve great outputs, this effect will only worsen.
It is critical that the AI tracking tool you choose enables you to understand the experience and information your ICP receives when they talk to AI about your problem space.
How Demand-Genius solves this.
Off the back of our research into the impact conversational context has on AI responses, we have extended our product to give every customer control over the specific tuning of their AI visibility tracking. Better yet, we recognise that many brands sell to multiple segments, ICPs and personas, each of which will engage with AI products specifically tuned to their context.
Everything in Demand-Genius’ visibility tracking now operates on a segment-by-segment basis.
- Configure an ICP segment: Business size, industry, requirements, key stakeholders.
- Configure the intent clusters you want to track within that segment.
- See how your brand shows up in key moments for your buyers in priority segments.
(An intent cluster is a list of prompts, but clustered around a specific buyer job-to-be-done or category entry point. This shows you if you are visible in a specific “moment” that matters.)
Want this applied to your brand?
Book a free audit. Real analysis of your AI position, no obligation.