Prompts are the questions and requests people type into ChatGPT, Gemini, Claude, or Perplexity, for example, “Which providers can you recommend for maintaining our production equipment in Bavaria?” Every AI answer starts with a prompt, including every answer in which your business is mentioned or left out.
What a prompt contains
A prompt can be just a few words, or it can be phrased like a conversation, in full sentences and with follow-up questions. It can also include details about the person’s situation: location, company size, budget, or specific requirements. Such details narrow down what the answer is about and therefore which providers are relevant. So an offering can appear in the answer to “Which provider is the cheapest?” and be missing from the answer to “Which provider is the best fit for a company with 200 employees?”
Yet even the same question in different words, or an identical prompt asked again, can get a different answer. This is known as answer variability.
What reaches the model besides the prompt
What you type is not everything that reaches the software that writes the answer, the large language model. Applications built on a language model can place their own instructions ahead of the prompt, for example on how the model should respond. Such instructions are called a system prompt. Technically, they are part of the model’s input, but in everyday use, “prompt” usually means only what a person types.
When an AI system retrieves content from the web for its answer, the language model receives that content along with the prompt, a method known as retrieval-augmented generation (RAG). To retrieve it, the AI system does not necessarily use the prompt’s exact wording but writes its own queries, sometimes several on individual aspects of the question; spreading a question out into such subqueries is called query fan-out.
Prompts in GEO
Because the answer depends on what is asked, and can vary even for the same prompt, GEO measurements of AI visibility use a fixed prompt set that reflects typical questions from the target audience. In prompt tracking, each of these prompts is asked repeatedly; metrics such as mention rate are derived from the results.
Good prompts for such a measurement are based on the questions customers actually ask. In GEOLYX, businesses enter the prompts that matter to their offering alongside their brand, domain, and competitors.