Prompt
A prompt is the input you give a language model — ChatGPT, Claude, Gemini or another — to tell it what you want. Usually it's a sentence or two: a question, a task, an instruction, sometimes with examples or constraints attached. The model reads the prompt together with whatever context it's been given and generates a response token by token. How you phrase it shapes what you get back, which is why the same request worded two different ways can produce two very different answers.
Why the prompt matters for AI visibility
In generative engine optimization, the prompt is the unit you measure with. People don't type keywords into ChatGPT or Gemini the way they used to type them into Google — they ask full questions: "Which accounting software works for a small retail business?" Whatever brands the model names in response to prompts like that are the ones showing up in the buying decision. If you want to know your AI visibility, you write a representative set of realistic prompts — the actual questions your customers would ask — and run them repeatedly to see whether and how often you're named. This matters because AI citation doesn't track search rankings the way you'd expect: one Ahrefs analysis found only 6-8% of URLs cited by ChatGPT overlap with a page's Google top-10 ranking for the same query, and roughly 80% of ChatGPT-cited URLs don't appear in Google's top 100 at all. Ranking well and being prompted well are different problems, and the prompt is how you test the second one.
How a prompt works technically
A language model doesn't read your prompt the way a person does. It splits the text into tokens — word pieces roughly a few characters long — and then predicts, one token at a time, the most probable continuation given everything it has seen so far. Your prompt, any examples you supply, the conversation history, and an often-invisible system prompt set by the developer all sit together inside the context window, the fixed amount of text the model can attend to at once. The more precisely you specify the task, the audience and the format, the less the model has to guess. Leave that out and it fills the gaps with assumptions, which is where vague or invented answers come from. Structuring a prompt deliberately — giving it a role, a goal, constraints, examples — is the practice known as prompt engineering, and it reliably produces more useful output than a one-line request.
Common mistakes
The most common failure is too little context: "write me a description" doesn't say of what, for whom, or how long. A close second is stacking several distinct tasks into one prompt, which tends to produce a half-finished answer to all of them rather than a complete answer to one. Padding a prompt with pleasantries or hedging doesn't help either — it just spends context window on nothing. A more consequential mistake is trusting the output uncritically: language models hallucinate, stating invented facts with the same confidence as real ones, so anything that matters needs an independent check. Columbia's Tow Center tested this directly — across 1,600 queries asking AI search tools to identify a news article's source, headline, date and URL, more than 60% of responses were wrong across every tool tested, ChatGPT alone missing 134 of 200. And for visibility measurement specifically, one prompt tells you nothing: phrasing and outputs both vary, so you need many prompt variants across several runs before the pattern is real.
Relation to AI recommendations
Every AI recommendation starts as a response to a prompt. Whether a model names your business when asked "who makes sustainable office chairs?" depends on how visible and citable your content already is — the prompt itself doesn't create that visibility, it just surfaces it. That's also why the fix rarely lives in the prompt: an Ahrefs study of roughly 75,000 brands found that how often a brand gets mentioned across the web correlates with AI citation rate at about 0.664, nearly three times the correlation for backlinks (about 0.218). For your own monitoring, the practical approach is to write down the prompts your customers would realistically ask, run them on a schedule across the AI systems that matter, and track your mention rate and share of voice against competitors. That turns the prompt from a one-off input into a recurring measurement of where you stand in generative search.
Example
A tax advisory in Leipzig wants to know whether AI assistants recommend the firm. Instead of typing "tax advisor Leipzig", the owner writes prompts the way an actual prospect would ask them: "I run a small online shop out of Leipzig and need a tax advisor who understands e-commerce and quarterly VAT — who should I look at?" She runs that prompt, plus nine phrasing variants, through ChatGPT, Perplexity and Gemini over the course of a week and logs whether her firm is named. She comes back with a number she can act on: named in four of the ten runs, a baseline she can now try to move.
Common questions
What's the difference between a prompt and prompt engineering?
A prompt is the specific input you send to the model. Prompt engineering is the discipline of designing prompts deliberately — with a role, examples, constraints and a clear output format — so the result is reliably good rather than left to chance.
How many prompts do I need to measure AI visibility?
More than one. A single prompt's answer can swing run to run, so you need several realistic phrasing variants per topic, repeated across multiple runs and ideally across more than one AI system, before the pattern reflects reality rather than noise.