How to Build a Prompt Set for AI Prompt Tracking (With a Template)

AI prompt tracking starts with a prompt set: a fixed list of buyer questions, each tagged and asked the same way on every run. It works like a survey. If you change the questions each time, you can't compare the results.

The list you choose decides what you measure, so it's the one part you can't fix later with a better chart.

People don't phrase one need the same way, either. SparkToro and Gumshoe asked 142 people to write a prompt for one intent, and the average similarity between those prompts was 0.081. In plain terms, the wording barely overlapped.

The steps are short: pick how many prompts, find real buyer words, tag each prompt and keep it all in a template.

Key takeaways

• A prompt set is a fixed list of buyer questions, asked the same way every run.
• Track at least 20 prompts before you read a percentage, and expect a wide range at that size.
• Tag every prompt by funnel stage, buyer type and whether it names a brand.
• Change no more than a quarter of the set at once, or you start a new baseline.

What is AI prompt tracking?

AI prompt tracking means asking the same buyer questions on AI engines on a schedule, and noting which brands each answer names. The questions are your prompt set. Each run asks every prompt on every engine, so the answers in a run are prompts times engines.

On the Firemark pricing page, checked on October 22, 2026, the Lite plan tracks 25 prompts on three engines. That's up to 75 answers per run. A bigger set means more answers to read, not a different method.

A prompt set isn't a keyword list. Keywords are short and steady. Prompts are full questions with details, like a buyer's real words, and the set only works if it stays the same from run to run.

That fixed wording lets you tell a real change from a wording change. A trend is several runs you can compare, and Firemark calls a movement only after three of them and a change of 5 points or more.

How many prompts do you need?

You need at least 20 prompts before a percentage means anything, and more for a tighter range. Firemark shows a percentage only from 20 prompts up, because a share on fewer prompts jumps around too much to read.

Say your brand is named in 40% of answers. On 25 prompts, the 95% range runs from 23% to 59%, using the Wilson interval. On 100 prompts it runs from 31% to 50%, on 200 from 33% to 47% and on 500 from 36% to 44%.

It's like a poll. Ask 25 people and the result is rough, and ask 500 and it's sharper.

Those ranges come from Firemark's method page. They're math on one engine's base, not data from a real brand. More prompts give a tighter range, but each one costs time to set up and read.

Start with 25 to 40 prompts across the funnel, then add more once you see which questions separate brands.

The cost adds up across engines. Each prompt gets one answer per engine per run, so 500 prompts on five engines is up to 2,500 answers every run. That's the top plan on Firemark's pricing page, and it's far from a first set.

Where do the prompts come from?

Strong prompts come from the words buyers already use, not from a brainstorm. Four sources cover most of what you need:

• Sales and discovery calls: ask new prospects what they typed into an AI tool, and write down the exact words.
• Support tickets and reviews: customers describe a problem before they know your category name.
• Search Console: export queries from the Performance report and keep the ones phrased as questions.
• Forums and People Also Ask: both show how people describe a problem in their own words.

Turn each phrase into a full question and keep the details you hear, like team size and budget. A note that says "we couldn't find a way to chase late invoices" becomes "how do small agencies chase late invoices without annoying clients".

Record the source of every prompt. If a prompt has none, mark it as brainstormed and review it first when you trim the set.

Write two or three phrasings for each intent. People with the same goal write prompts that barely look alike, as Why Everyone Prompts AI Differently shows. One phrasing can't stand in for the whole intent.

Do my prompts need to be the exact words a customer would use?

No. A prompt doesn't have to match a customer's exact words, because AI tools work from what a question means, not only the words in it.

ChatGPT can also rewrite your wording into search queries of its own, which OpenAI's help page describes. It may even send follow-up searches that are more specific after it reads the first results.

That means AI tools don't treat your prompt like an exact-match keyword in paid search. They behave more like broad match: they work from what you mean, not only the words you typed.

Google calls its version query fan-out. AI Overviews and AI Mode may send several related searches across subtopics, then build one answer from the results.

So a prompt that isn't a perfect match doesn't automatically lose you data. A brand can be named for a question worded a little differently from yours, because the engine works from the intent. SparkToro's study points the same way: people wrote prompts that barely looked alike for one goal, and the AI tools often returned similar groups of brands.

Wording can still shift the list a little. That's why you track the intent, not the exact string, with two or three phrasings for each one.

How should you tag each prompt?

Tags work like labels on file folders: they let you sort the answers later. Tag each prompt with three things: its funnel stage, its buyer type and whether it names a brand. Firemark gives every prompt a funnel stage, so you can see where in the buying journey an engine names you.

Funnel stage has three values. Research prompts name no product, like how to track client invoices. Compare prompts build a shortlist, and decide prompts sit close to a purchase, like a head-to-head or a price question.

Buyer type is who is asking, like an owner, an operations lead or a finance lead. It shows whether you're named when a finance lead asks but missing when an owner asks. "Names a brand" has three values: none, yours, or a rival.

Keep prompts that name your brand apart from category prompts. A prompt with your own name in it almost guarantees a mention. Mixing the two inflates a share that should show real competition.

What should my DIY prompt set template look like?

A template has one job: it keeps your prompts, your answers and your changes in one place, so every run asks the same questions the same way. That's what lets you compare one run with the next, and tell a real change from a wording change.

Copy this table into a sheet, one row per prompt. Fill the last column only when a prompt names a brand. The examples are for a sample invoicing tool for small agencies.

Get the ready-made version

Want it ready-made? Download the AI Prompt Set Template (.xlsx). It's a spreadsheet with sample data to overwrite. Here's what's inside:

• Prompt set: one row per prompt, with the ID, prompt, intent, funnel stage, buyer type and whether it names a brand, plus where the prompt came from and the date you added it, so the set has a history. Drop-downs keep the tags consistent, and duplicate wording gets flagged. An even split across the three stages is a good start, so a set of 30 has about ten prompts per stage.
• Answers log: one row for each prompt, engine and run: prompt ID, engine, run date, named and linked. Count named and linked separately, because a link alone isn't a mention.
• Read by stage: fills in on its own from the first two tabs. It shows how often you're named and linked in research, compare and decide prompts, for every prompt and for category prompts only, and it warns you while the set is under 20 prompts.
• Change log: dates every change and tells you when you've changed more than a quarter of the set, which starts a new baseline.
• How to use and Lists: the steps, and the definitions behind each tag.

Once the first run is in, read the set by stage before you read it as a whole. A brand named in most compare prompts but missing from research prompts has an awareness gap, and the reverse points to a shortlisting problem.

A flat average hides both, which is why the tags matter.

The demo report shows what a tagged set looks like once it has run, using sample data.

Is a DIY template worth the time?

A DIY template works, but it's slow. Every run means asking each prompt on each engine, reading the answer and logging who was named and linked. On a 25-prompt set across three engines, that's up to 75 answers a run, and a trend only shows after several runs.

That's a lot of hours for work a tool can do for you. Automating the collection is usually the smarter use of your time, because you spend it reading the results and fixing the gaps instead of copying answers into cells.

That's the part Firemark takes off your hands. It runs your prompt set across the engines and records which brands each answer names, and it gives every prompt a funnel stage so you can see where in the buying journey you're named. The entry plan covers 25 prompts, enough for a first set, and a new account gets one free run of 10 prompts on 4 engines, so you can see a tagged set in a report before you spend anything.

What should you avoid?

Weak prompt sets fail in four common ways, and each is easy to avoid.

• Your brand name in a category prompt, which inflates the count.
• Near-duplicate prompts, which count one intent several times.
• One "ideal" prompt, because buyers don't phrase a need once.
• Constant rewrites, which break the trend you're trying to read.

Firemark treats a change to more than a quarter of the prompts as a new comparison baseline, because the runs after it no longer measure the same thing. It's like changing the test questions halfway through the school year. Date every change and swap prompts in small batches.

Review the set each quarter, and drop a prompt that never names any brand, because it can't tell you apart from anyone. Definition questions like "what is invoicing" usually land in that group. An answer explains a concept and recommends nobody.

Can an AI draft the set for you?

An AI assistant can draft a prompt set, but a person has to confirm it before anything runs. With the Firemark MCP server, Claude or ChatGPT can interview you in five steps (brand, competitors, buyer types, offering and review) and save a draft. A member of the account approves or discards the draft in the app, and nothing runs or is charged before that.

Read an AI-drafted set before you confirm it. Check that category prompts don't contain your brand name, that no prompt names a rival you don't track, and that each prompt sounds like something a buyer would type.

You can also generate prompts in the app, import your own from a CSV or write them by hand. A sheet is enough for a first set, and a tool earns its place once you track more engines or need history.

Start with 25 prompts and the template above. Firemark plans start at 25 prompts, and a new account gets one free run of 10 prompts on 4 engines.