Most teams start prompt tracking by typing their brand name into ChatGPT and screenshotting the answer. That gives you a data point, not a dataset. The screenshot changes tomorrow, you cannot compare it to anything, and it tells you nothing about the questions buyers actually ask before they know your name exists.
A prompt set is a different thing. It is a fixed list of questions you commit to running on a schedule, across engines, with the same wording every time. The method below builds that list from four intents, keeps it free of brand names, and holds it stable long enough to produce trends instead of anecdotes.
Start with four intents, not a keyword list
Buyers move through recognizable stages, and each stage produces a different kind of question. Sort your prompts into four buckets.
Informational
The person is learning the problem space and does not know the vendors yet.
- "How do you reduce customer churn in a subscription business?"
- "What does a compliance review involve for a new payment processor?"
These prompts are where you find out whether AI engines cite your category content at all. If you are absent here, you will rarely appear later either — the engine has no reason to have learned your name.
Commercial
The person is evaluating approaches or vendors, still without naming a brand.
- "What should a mid-size company look for in a customer support platform?"
- "How much does it cost to migrate a data warehouse?"
Commercial prompts are usually the most valuable bucket for pipeline, and the hardest to win, because the answer is often a list of vendors. That is exactly why you track them.
Local
Geography changes the answer, so geography belongs in the prompt.
- "Best bookkeeping services in Austin for a small retail business"
- "IT support companies near Manchester for a 50-person office"
If you serve specific cities or regions, run one local prompt per market you care about. Do not try to cover forty cities at once; pick the markets that carry revenue.
Comparison
The person is choosing between named options. This is the one bucket where mentioning other companies is fine — and necessary.
- "Compare HubSpot and Zoho for a 10-person sales team"
- "Alternatives to [a well-known competitor] for enterprise security teams"
Comparison prompts are noisy. Answers shift as vendors publish new content, and engines disagree with each other. Treat movement here as a signal to investigate, not a score to report.
Keep your brand name out of the prompts
This is the rule people break first. If your prompt is "Is Acme any good?", the engine will produce something about Acme, and you will learn almost nothing. You already know you exist. The question is whether you show up when the buyer has not heard of you.
The exception is a small, deliberate brand set: two or three prompts that check what an engine says when someone asks about you directly. Those catch factual errors — wrong pricing, stale leadership, a discontinued product described as current. Keep them in a separate group and do not mix them into your visibility numbers, because they measure something different.
One more caution: prompts that name competitors tend to produce answers dominated by whichever vendor has the most published comparison content. That is real information, but it is a measure of that vendor's content operation as much as your own standing.
Aim for 20 to 50 prompts, weighted toward the middle
Twenty is the floor for anything you would call a trend. Fifty is roughly where a manual review stops being sustainable and where each additional prompt adds less than the last.
A workable split for a B2B company:
| Intent | Share of set | Example count (of 30) |
|---|---|---|
| Informational | 30% | 9 |
| Commercial | 35% | 11 |
| Local | 15% | 4 |
| Comparison | 20% | 6 |
The weighting matters. Informational prompts are easy to write and easy to win, so teams over-index on them and end up with a dashboard that looks healthy while commercial queries — where the buying decision happens — stay empty. If your product is local by nature, flip the local and informational shares.
A starter template
Use this to draft your first set. Fill the brackets with your category, not your brand.
| # | Intent | Prompt pattern | Your version |
|---|---|---|---|
| 1 | Informational | How do you [core job to be done]? | |
| 2 | Informational | What are common mistakes when [task]? | |
| 3 | Informational | What does [process] involve step by step? | |
| 4 | Commercial | What should a [company size] [role] look for in a [category]? | |
| 5 | Commercial | How much does [category] typically cost for [segment]? | |
| 6 | Commercial | Which [category] tools work well for [specific constraint]? | |
| 7 | Local | Best [category] in [city] for [segment] | |
| 8 | Local | [Category] providers near [city] that handle [need] | |
| 9 | Comparison | Compare [option A] and [option B] for [use case] | |
| 10 | Comparison | Alternatives to [well-known competitor] for [segment] |
Write each prompt the way a person would type it, not the way you would write a search query. Full sentences and specific constraints produce more useful answers than two-word fragments.
Stability is the whole point
A prompt set is only useful if it stays still. Every edit resets your baseline, and a set that changes weekly produces a chart that measures your own edits.
Practical rules:
- Freeze the wording. "Best CRM for startups" and "top CRM for small companies" are different prompts. Pick one and keep it.
- Record the engine and date. An answer from one assistant on one day is a snapshot, not a trend.
- Note model updates. When an engine ships a major change, write it down. Movement after that date has an obvious candidate explanation.
- Do not tune prompts to flatter you. If you rewrite a prompt because you dislike the answer, you are no longer measuring.
Run the set on a fixed cadence — weekly is common, monthly is defensible for slow-moving categories. More frequent runs mostly produce noise.
Retire prompts deliberately
Prompt sets do drift, and that is fine, but retirement should be a decision, not an accident. Replace a prompt when one of these is true:
- The product, service or market it describes no longer exists in your business.
- The wording has stopped matching how buyers actually talk, and you can point to real language that replaced it.
- The prompt consistently returns answers that are unrelated to your category, which usually means the phrasing is ambiguous.
When you swap a prompt out, keep the old one in an archive rather than deleting it. You lose the trend line either way, but the archive lets you explain the gap later. And change one or two prompts at a time, not ten — a wholesale rewrite makes the whole set unreadable for a quarter.
A note on what you can and cannot control
Being cited in an AI answer depends on crawler access and on whether the engine has indexed and chosen your content. The major vendors document which bots do what: OpenAI's OAI-SearchBot surfaces sites in ChatGPT search, Anthropic's Claude-SearchBot indexes content for search results, and Perplexity's PerplexityBot surfaces and links sites in Perplexity results. Google-Extended controls training and grounding for Gemini models but does not affect Google Search inclusion, and AI Overviews are governed by normal Googlebot crawling.
What none of that gives you is a way to force a mention. There is no documented mechanism for buying or guaranteeing placement in an AI answer. Prompt tracking tells you where you currently appear and where you do not. The work of changing that is content and access, not prompt selection.
Pick your first ten this week
Write ten prompts — three informational, four commercial, two local if geography matters, one comparison — using the template above. Run them across the engines you care about, save the raw answers with the date, and do not edit the wording for thirty days. If you want a structured way to think about what those answers do and do not prove, How to Measure AI Visibility Without Fooling Yourself covers the measurement side. PromptRanking tracks whether ChatGPT, Gemini and Perplexity recommend a brand across a set like this, which is the point at which a list of questions becomes a trend you can act on.