We get asked this question more than almost any other at the moment, and nobody can give you a satisfying answer to it, however good their dashboard looks.
The real answer is you can’t.
Nobody hands over the prompts
ChatGPT, Gemini, Claude and Perplexity don’t share what users type. That applies to ad platforms too: ChatGPT’s ad model matches ads to the meaning of a conversation (semantic matching), so an advertiser learns which themes their ad appeared against but never sees the prompt a person wrote. It’s a sensible privacy stance, and it means there’s no equivalent of the Google Search Console query report for AI.
Prompts don’t behave like search terms either. A search is three or four words. A prompt is often a paragraph, full of context about the person’s role, budget, tech stack and deadline, and it usually runs as a conversation of five or six follow-up questions. Even if you could see every prompt, you’d rarely find two the same.
What the tools are showing you
Tools like Semrush’s AI visibility reports are useful, but it helps to know what they’re measuring. They infer likely prompts by starting from your content, your keywords and your existing SEO signals, then running synthetic questions through the models to see whether you get mentioned. That tells you how visible you are for questions the tool thinks people might ask. It can’t tell you what your buyers actually ask, and since it starts from your current content, it tends to confirm the topics you already cover and miss the ones you don’t.
Start with the people who hear the questions
The most reliable source of real customer questions already works for you. Your salespeople hear buyers describe their problems in their own words every week. Your technical team fields the detailed questions that come up once the buyer is serious, plus the problems that come in after the sale. Between them, they know how your customers think far better than any tool can estimate (we’ve written before about why your technical team might be your best marketing asset).
Get sales and technical people in a room and plot the customer journey together, writing down the questions a buyer would ask an LLM at each stage, in the buyer’s language rather than yours.
Map the questions to four stages
1. Awareness: problem-led category entry points
The buyer knows they have a problem but may not know your category exists. Their prompts describe symptoms: “Why do our field engineers keep getting sent to jobs without the right parts?” If your content only talks about your solution, you won’t show up here.
2. Research: solution-focused
The buyer now knows what type of solution they need and wants to understand the options. “What’s the difference between a radar and a Lidar platform?” “What should a mid-sized manufacturer expect to pay?” Clear, honest explanations win here.
3. Consideration: evaluation
The shortlist stage. Prompts get specific and comparative: “Which field service platforms integrate with SAP?” “What are the common problems when rolling out X?” Your technical team knows the objections, integration questions and implementation worries that decide deals.
4. Customer success: FAQs and how-to support
This stage gets forgotten, and it matters more than most. Existing customers ask LLMs how to use your product, fix an error or set up a feature. If the model gives a poor or outdated answer, or points them to a competitor’s documentation, you risk the renewal. Detailed FAQs, how-to guides and troubleshooting content keep customers loyal and give the models accurate material to draw on.
Turn the map into content
Once you’ve mapped the questions, you have a content plan built on real buyer intent rather than guesswork. Check each stage against what you already publish, fill the gaps and use tools like Semrush to track whether your visibility improves. The tools work best as a way of measuring progress, with your own people deciding what goes in the plan and what is tracked.
You’ll never see your customers’ prompts. What you can do is know their questions well enough that it doesn’t matter.