How to Build an AI Prompt Set (+ Free Template)

Most first prompt sets are keyword lists with question marks bolted on. They produce clean-looking charts that describe searches nobody makes. This lesson shows you how to build a 30-prompt set from the questions your buyers actually ask, how to tag it so the results mean something, and how to run it enough times that a single lucky answer doesn't fool you.

In this lesson

• The problem: a prompt list that lies to you
• Why keyword lists make bad prompts
• The four intents and the shape of a 30-prompt set
• Where real buyer questions hide
• Writing prompts that sound like people
• Worked example: Ledgerline's first 30 prompts
• Running prompts: why one answer isn't a result
• From MentionX data: how much answers move
• Tagging, versioning and storing the set
• The mistakes we see most
• Exercise: build your first 30
• Quiz

Throughout this lesson, anything marked "Our view" is our opinion from running prompt sets, not documented engine behaviour. Ledgerline, the company in the worked example, is fictional.

The problem: a prompt list that lies to you

Picture a marketing lead at a mid-sized software company. Her CEO has read that buyers now ask ChatGPT for vendor shortlists, and wants to know whether the company shows up. She opens the SEO keyword tracker, copies the top 20 keywords, turns each into a question ("What is accounts payable software?"), pastes them into ChatGPT once each, and counts the brand mentions. Six out of 20. She reports 30% AI visibility.

Almost everything about that number is wrong, and none of it is her fault. She's never been shown how to do it properly.

• The questions aren't real. "What is accounts payable software?" is a keyword in a costume. A finance director with a real problem asks something closer to "We're a 200-person company drowning in supplier invoices in Excel. What should we look at?"
• The mix is skewed. Keyword lists lean towards high-volume definition terms, so the set is mostly early-stage learning questions, where vendor names rarely appear at all. She's measuring the part of the journey where she's least likely to show up, and missing the shortlist questions where it matters.
• One run is a coin flip. Ask the same question tomorrow and the answer may name different vendors. Six out of 20 on one afternoon could be three or ten next week.

So the prompt set is the foundation of everything else in this course, and it builds on the quick check you did in Lesson 4: Your first manual visibility check. Your baseline scan in Lesson 6, your results in Lesson 7 and your 30-day plan in Lesson 8 all inherit its quality. Get it wrong and you'll spend a quarter fixing the wrong pages.

Why keyword lists make bad prompts

Search keywords and AI prompts are written by the same people, but in different moods. A search box trains people to compress: "ap automation netsuite". A chat box invites them to explain: who they are, what they've tried, what they can spend, what's annoying them.

That extra context changes the answer. An assistant asked about "AP automation" might list the biggest names in the category. Asked about AP automation "for a 200-person company on NetSuite that pays suppliers in three currencies", it has to narrow down, and that narrowing is exactly where a smaller, more specialised vendor can win.

The engines also do something with your words before answering. (If you skipped it, Lesson 2: How AI engines build answers covers the basics.) OpenAI's help article on ChatGPT search says that when it uses search partners, ChatGPT "typically rewrites your query into one or more targeted queries" and may send further, more specific queries after reviewing the first results (OpenAI Help: Searching the web with ChatGPT). The same article says it may use approximate location from your IP address and, if memory is on, saved memories when rewriting.

Our view: that has two consequences for how you write prompts. First, the constraints in a prompt (team size, software stack, country) probably carry into the searches the engine runs, so constraint-rich prompts tell you which pages the engine finds for buyers like yours. Second, who is asking, and from where, can change the answer. We'll come back to that when we talk about running prompts.

The four intents and the shape of a 30-prompt set

We sort every prompt into one of four intents. Three of them follow the buying journey; the fourth checks whether the engines describe you accurately.

Why 30? Our view: it's the smallest set where each intent has enough prompts to show a pattern, and it's still small enough to run several times a month by hand if you don't have a tool. Going bigger before you have someone to act on the results just produces more numbers nobody reads. Lesson 6 covers how many runs per prompt you need; for now, keep the set tight.

Why weight compare prompts most? Because that's where a vendor name is the answer. In a learn prompt the assistant often explains a concept and names nobody. In a compare prompt it almost always produces a list, and being on that list or missing from it is the single most useful thing you can measure early on.

Brand-check prompts are different. They aren't about visibility; they're about accuracy. Lesson 3 explains why a mention, a citation and a recommendation are worth tracking separately. If an assistant says you don't integrate with a system you've supported for years, that's a fix you can make this week, and it's often the fastest win in the whole programme.

Where real buyer questions hide

You need prompts written in your buyers' words, not yours. Here are the five sources we use, roughly in order of value, with the tactics we've found most useful for each.

1. Sales calls and demo notes

Your sales team hears the real questions every day, unfiltered. Ask each rep for the three questions prospects ask in the first ten minutes of a discovery call, word for word. If you record calls, search transcripts for phrases like "we're currently using", "we looked at" and "the problem is". Those phrases reveal the competitors buyers compare you with and the constraints they care about.

2. Search Console: long, question-shaped queries

Search Console's Performance report shows the queries your site appeared for (Google: Performance report overview). Most are short. The long ones are gold, because people who type nine words into Google are already writing prompts.

The report supports custom regular-expression filters on queries, using RE2 syntax, with partial matching and case-insensitive matching by default (Google: Advanced filtering and comparison). Google's own Search Central blog suggests filtering for question words such as what|how|when|why (Google Search Central Blog, 2021).

3. Reddit and other communities

Communities are where people write out their whole situation, because they want a human to answer. Search Reddit for your category plus words like "recommend", "switching from" or "anyone using". Read the opening post, not just the title: the second paragraph usually holds the constraints that matter ("we're on NetSuite", "budget is tight", "auditors want approval trails").

This matters beyond phrasing. Reddit has data partnerships with both Google and OpenAI (Google, Reddit and OpenAI), so threads like these are part of the material assistants can draw on.

4. People Also Ask

Google's People Also Ask boxes show related questions for a search, and clicking one usually loads more. Search your top five category terms, expand the boxes a few times, and write down every question that a buyer (rather than a student) would ask. It's a fast way to fill gaps in your learn and buy intents, especially questions about cost, implementation time and risk.

5. The assistants themselves

The assistants are a source too, used carefully. Ask ChatGPT, Gemini or Perplexity to play your buyer:

"I'm the finance director at a 200-person company. We process around 1,500 supplier invoices a month, mostly by email and Excel. List the 15 questions I'd realistically ask an AI assistant while choosing software to fix this, in my own words, from first noticing the problem to signing a contract."

Treat assistant-generated prompts as drafts. Our view: no more than a quarter of your set should come from this source, and every one should be checked against something a human actually said. Otherwise you risk measuring how the model talks to itself.

Writing prompts that sound like people

Before a prompt goes into the set, run it through four quick checks:

1. Read it aloud. Would a busy person really type this? If it sounds like a heading, rewrite it.
2. Count the constraints. Most compare and buy prompts should include at least one: company size, industry, existing software, country or budget.
3. One question per prompt. "What's the best tool and how much does it cost?" produces a muddled answer you can't score.
4. No brand in non-branded prompts. Your name belongs only in brand-check prompts and in direct "X vs Y" prompts where buyers really do name you.

Worked example: Ledgerline's first 30 prompts

Illustrative example. Ledgerline is a fictional company, and the prompts, sources and volumes below are invented to show the method. No scan results are shown because none were run.

Ledgerline sells accounts payable automation to mid-market finance teams, mostly companies with 100 to 1,000 staff, running NetSuite or Microsoft Dynamics, in the UK and the US. Its marketing manager spent one afternoon gathering material: notes from 12 recent sales calls, a Search Console export filtered with the regex patterns above, eight Reddit threads, People Also Ask boxes for five category terms, and a "play the buyer" session in ChatGPT.

Here's the set she ended up with. Competitor names are placeholders for the real competitors her sales team named most often.

A few things are worth noticing.

• Sources are spread out. No source supplies more than about a third of the set. If every prompt had come from Search Console, the set would reflect people who already found Ledgerline's site.
• The status-quo prompts (L01, L07) don't mention the category. Our view is that these are among the most revealing prompts in the set. If the assistant's answer to "Excel is falling apart" never reaches AP software, buyers are being steered somewhere else before they've even heard of the category.
• Only six prompts name Ledgerline (C08 and R01 to R05). The other 24 measure whether Ledgerline turns up when nobody asks for it, which is the real question.
• IDs are stable and readable. "C04" means the same prompt in every report for as long as this version of the set exists.

Running prompts: why one answer isn't a result

Here's the part most guides skip. Language models don't look up a fixed answer; they generate one, choosing each piece of text with an element of chance. Google's documentation for Gemini models describes temperature as controlling "the degree of randomness in token selection", and says that even at a temperature of 0, "a small amount of variation is still possible" (Google Cloud: Experiment with parameter values). The same page says a fixed seed is only a "best effort" and that "deterministic output isn't guaranteed". In consumer chat apps you don't control any of these settings.

Add web search on top, where the engine may rewrite your question into different searches and read different pages, plus location and memory effects, and you get answers that shift from run to run.

So stop thinking of a prompt as having an answer. Think of it as having a distribution of answers. Your brand doesn't "appear" for C01; it appears in some share of the answers to C01. That share is the number you're after.

Our view: the variance isn't noise to complain about; it's information. A brand that appears in nine of ten runs has a stable position. A brand that appears in two of ten is on the edge of the shortlist, and that's often where a little work moves the most.

How many runs? Lesson 6: Running a baseline scan covers sampling properly. As a starting rule, we run each prompt at least five times per engine before drawing any conclusion, and we never report a single run as a result.

From MentionX data: how much answers move

We checked this against our own scan data rather than take it on trust.

Method. We took every case in the MentionX scan database where the exact same prompt was run, for the same brand, on the same engine, at least three times. That gave 28 prompt-and-engine pairs and 643 completed runs, across two engines (OpenAI's Responses API with web search, and Gemini with Google Search grounding), between 4 September and 19 October 2026. Most pairs were run 23 times, once per scan. For each run we recorded whether the tracked brand was mentioned in the answer. All 28 pairs belong to one brand, which we've left anonymous.

What we found.

Among the 11 mixed pairs, the brand's mention rate ranged from 2 of 23 runs to 22 of 23. Five of them landed between 7 and 11 out of 23, so a single run of any of those prompts was close to a coin flip. In 19 of the 28 pairs, the number of brands the extraction picked up also changed between runs.

How to read it. Repeated runs were spread over about six weeks, so some of the movement is the engines and the web changing over time, not only the randomness of a single answer. This is one brand on two engines, so treat it as a demonstration, not a benchmark. It still makes the lesson's point: for about four in ten of these prompts, one run would have given you the wrong picture.

Tagging, versioning and storing the set

Keep the set in one shared sheet. Our free template has the columns you need to start: ID, Intent, Prompt, Source, Tags, Runs, Cited sources, Notes. Download the blank 30-row CSV and the Ledgerline example CSV, then import them into Google Sheets or Excel. As the set grows, add market, version, status, date_added.

Rules we follow:

• Freeze the set for at least a quarter. Trends only mean something if the questions stay the same.
• Version every change. Adding prompts makes v1.1; editing wording makes v2. Only compare runs within the same version.
• Retire, don't delete. Mark old prompts as retired so past reports still make sense.
• Change for good reasons only. A new product, a new market, or evidence from sales and support that buyers ask something you aren't tracking. Never remove a prompt because you look bad in it.

The mistakes we see most

1. Measuring definitions. Sets that are mostly "what is X" questions, where no vendor gets named. They look safe and tell you almost nothing.
2. One source only. All keyword tool, or all assistant-generated. Each source has blind spots; mix them.
3. Leading prompts. "Why is Ledgerline the best AP tool?" Of course the assistant plays along. Brand-check prompts should be neutral.
4. Near-duplicates. Five versions of "best AP software" make your results look more stable than they are and waste runs.
5. Ignoring the competitor buyers really compare you with. Ask sales; it's rarely the one marketing assumes.
6. Running once and reporting. The opening story of this lesson. Don't be the 30% chart.
7. Quietly editing prompts mid-quarter. It breaks every comparison you'll want to make later.

Exercise: build your first 30

Set aside about two hours.

1. Gather raw material (60 minutes). Get three questions per rep from sales. Run the three Search Console regex filters above and export. Read five Reddit threads in your category. Expand People Also Ask for your top three category terms. Do one "play the buyer" session in an assistant, with the follow-up move.
2. Draft 40 candidates (30 minutes). Rewrite each in the buyer's voice. Run the four checks.
3. Cut to 30 (15 minutes). Aim for 8 learn, 10 compare, 7 buy and 5 brand-check prompts. Remove near-duplicates. Make sure no single source supplies more than about a third.
4. Review with someone from sales or support (15 minutes). Ask: would a buyer really say this? What's missing? Is anything biased towards us?
5. Save it as v1 in the template CSV with IDs, tags and sources, ready for Lesson 6.

Quiz

1. Why do long Search Console queries make good prompt seeds?

a) Google ranks them higher b) People who type long queries are already describing their situation, as they would in a chat c) They have the most search volume Answer: b

1. Which is a compare-intent prompt?

a) "What does accounts payable automation do?" b) "We've outgrown Competitor A and need multi-entity approvals. What should we look at?" c) "How much does Ledgerline cost?" Answer: b

1. You ran a prompt once and your brand appeared. What can you conclude?

a) You rank for that prompt b) Very little; answers vary between runs, so you need several runs to estimate how often you appear c) The engine prefers your site Answer: b

1. Why run key prompts in a logged-out window too?

a) Logged-out answers are always better b) Memory and history can influence answers, so a clean session shows what a new buyer is more likely to see c) Engines block logged-in testing Answer: b

1. What should you do when a prompt shows you in a bad light?

a) Delete it b) Rewrite it to include your brand c) Keep it; it's showing you exactly where to work Answer: c

Next: Lesson 6: Running a baseline scan