A good AI prompt has four parts. Once you can name them, prompt research for AI visibility stops being guesswork. Check every prompt you track against all four, and source each part from research rather than assumption.
Prompts behave differently from search queries. Semrush analysed more than 80 million lines of ChatGPT clickstream data from late 2024. Prompts without search averaged 23 words, against about four words with search switched on. People describe their situation in natural language, so short keyword lists miss the situation and the limits. Those are often the parts that decide whether a business gets named.
The Four Parts of an AI Prompt
When someone asks AI for a recommendation, they usually pack four things into one question:
“Recommend [what they want] an accountant [product or service] for a sole trader tradie who’s outgrown doing his own books [situation] on a fixed monthly fee [limits]”
- What they want: recommend, compare or find.
- The product or service: what they’re buying, in their words, not yours.
- Their situation: who they are and the problem they’re solving.
- Their limits: price, location, timing or a must-have feature.
Most businesses only cover the first two. Yet the situation and the limits are what separate one buyer from the next.
How to Build an AI Visibility AI Prompt, One Part at a Time
Take an accounting firm. A keyword-based list would start with “small business accountant”, which is a keyword rather than a prompt. “Recommend an accountant” is a bare prompt that gives AI little to work with. Add the situation, “for a sole trader tradie who’s outgrown doing his own books”, and AI has a reason to favour some firms over others. Then add the limit, “on a fixed monthly fee”, because customers really do ask it that way.
Part 1: What They Want
The request is usually recommend, compare or find. It shows up in People Also Ask, autocomplete, support tickets and sales calls. The usual mistake is tracking only “best accountant” style prompts. Customers also ask AI to compare two options or find someone nearby, and each request can produce a different answer.
Part 2: The Product or Service, in Their Words
This is what customers call the thing you sell. People Also Ask, Google autocomplete and Search Console show the words buyers use, and reviews show how they describe what they bought. Avoid your own product names. If your website says “integrated wealth solutions” but customers ask for “a financial adviser for retirement”, use their phrasing. That’s the question AI is actually being asked.
Part 3: Their Situation
This is who the customer is and what’s prompting the question. Customers often start from the problem rather than the product. A dairy farmer often asks why their cell count is rising, not which teat spray to buy.
Most businesses leave this part out, yet it’s what makes a prompt realistic. Reviews, support tickets and sales calls show who bought and why. There’s no reliable volume data for AI prompts, so a situation that keeps coming up there is a good sign it matters.
Part 4: Their Limits
Limits are budget, location and timing. Sales calls and support tickets show price and availability concerns, and autocomplete shows modifiers such as a town name or “near me”. Only add the limits customers actually type. Stacking invented ones, such as a suburb, a budget and a deadline in one prompt, narrows your tracking for no reason.
Why AI Visibility Tool Prompts Miss the Parts That Matter
Many AI visibility tools generate suggested prompts from a summary of your website, so the result reflects how your marketing describes you. The situation is usually missing, because website copy rarely captures why individual customers bought. Tracking that list mostly tells you how well AI repeats your own positioning back to you.
Building the list yourself from customer language takes about an hour:
- Pull questions from People Also Ask, autocomplete, Search Console and your reviews.
- Ask whoever answers your phones what customers ask them.
- Write 10 to 15 prompts and check each one against the four parts.
- Flag any part you can’t source as an assumption rather than inventing it.
Include the “Do I Even Need This” Questions
Include two or three prompts from before the customer has chosen a category. Pew Research looked at almost 69,000 Google searches made by US adults. An AI summary appeared on 8% of one or two word searches but 53% of searches of ten words or more. Longer, more situational questions are where Google’s AI summaries show up most.
In one recent ADMATIC trial, AI answered a “do it yourself or get help” question by recommending a professional, then named the business we were working with. These prompts won’t always name businesses, so check each one before locking it in:
- If the answer only cites government sites, such as the ATO or Inland Revenue, cover the topic on your website and leave it off the tracking list.
- If it names competitors or directories, keep it, because that’s an answer you can realistically get into.
Track the Same AI Visibility Prompts Every Month
Build up to 20 to 30 prompts across the customer journey and marketing funnel. Run them monthly in fresh chats while logged out, and record who gets named and which sites get cited. Keep the list stable for a quarter, because AI answers vary between runs and you need consistency to see real movement.
That consistency gives you evidence you can take to the business. Start with the hour and the spreadsheet. If you’d like a second pair of eyes, ADMATIC offers a free first look at how AI search treats your business.