A good AI visibility prompt set is a short, fixed list of questions your buyers would actually ask, grouped by intent: category, comparison, problem and brand. Most of the list should be questions that do not contain your name, because prompts that name you mostly measure whether the model knows you exist. Keep the set small enough that a person reads every answer, and change it on a schedule rather than whenever a result disappoints you.

What a prompt set is for

A tracked prompt is one question, run repeatedly, so you can watch how AI assistants answer it over time. Ahrefs describes tracked prompts as a focused way to monitor answers to specific questions, as distinct from its much larger Brand Radar dataset [1]. The set is your sample of the market. If the sample is skewed, every number or screenshot you pull from it is skewed too.

One answer is also a weak signal. Semrush notes that one prompt can produce different answers across runs and platforms, so a single answer is only a partial view [4]. That is the main reason to keep the set stable: you compare the same question over time, not one lucky or unlucky response. For the statistical side of this, see why single-run percentages mislead.

The four intent groups

Sort every candidate prompt into one of four groups. A prompt that fits none of them is probably too vague to track.

GroupWhat the buyer is doingWhat you learnExample pattern
CategoryLooking for a type of productWhether you are named when no brand is"best [category] for [segment]"
ComparisonChoosing between optionsWhether you enter the shortlist and how you are described next to rivals"[Competitor A] vs [Competitor B]", "alternatives to [Competitor A]"
ProblemDescribing a pain, not a productWhether you appear when the buyer has not yet named a category"how do we stop [problem]"
BrandChecking you specificallyWhether facts about you are correct"what is [Brand]", "does [Brand] integrate with [tool]"

Two notes on the table:

  • Comparison prompts need not name you. "Alternatives to [Competitor A]" tells you whether you are in the consideration set for people already evaluating a rival. Adding "[Brand] vs [Competitor A]" is fine, but keep it in the brand group for reporting.
  • Problem prompts are the hardest to win and the most informative. They show whether an assistant connects your product to the situations that lead people to buy.

Avoid branded bias

A prompt that contains your brand name will usually return your brand. Counting those results next to category and problem prompts inflates your visibility and hides the gaps.

Ways to keep the set honest:

  • Report groups separately. Never merge brand prompts into one headline figure with unbranded prompts. Brand prompts answer "is what the model says about us accurate?" Unbranded prompts answer "do we get found?"
  • Strip your name and product terms from category and problem prompts. Also avoid your own marketing phrases. Buyers use their own words, not your tagline.
  • Use competitor names deliberately. Naming a rival in a comparison prompt is fine because you are testing whether you get added, but note it in the sheet.
  • Do not tune prompts to a result. If you rewrite a prompt until your brand appears, you have built a prompt nobody else is asking.

Write prompts the way a buyer would type them: clear, specific, and in plain language. Include the context that changes the answer, such as company size, industry or required integration, and leave out filler.

Where to get the prompts

Start from real language, not from a brainstorm about your own positioning. General sources that work:

  • Sales call notes and the first questions prospects ask
  • Support tickets and onboarding questions
  • Search Console queries and the questions people ask in search suggestions (see the keyword research guide for method)
  • Community threads, review sites and competitor comparison pages
  • Your own pricing, integration and "vs" pages, which show what you already expect buyers to compare

If you use an assistant to draft candidates, treat its output as a rough list. Edit each prompt by hand, remove duplicates, and discard anything no buyer would say.

How many prompts to track

There is no correct number, and I would not trust a vendor or guide that gives one without explaining the constraint. The practical limits are budget and review time.

Budget. If you use a paid tracker, count checks, not prompts. Ahrefs defines one check as one prompt execution times one LLM times one location [1]. Its custom prompt tracking allowances are 150 checks a month on Lite, 300 on Standard and 600 on Advanced, and Starter does not include them [1]. Prompts added across several assistants or countries use up an allowance faster than the prompt count suggests. Check how your own tool counts scheduled refreshes before you decide.

Review time. Semrush recommends tracking buyer-decision prompts across AI platforms and reviewing them at least weekly to spot patterns in visibility, mentions, cited sources and prompt performance [4]. A set you cannot read each week will not get read. If the answers sit unread in a dashboard, the set is too big.

A workable approach, offered as our suggestion rather than a rule:

  1. Pick one product line or audience segment at a time.
  2. Write a handful of prompts per intent group.
  3. Run it for a few cycles and read every answer.
  4. Add prompts only where you find a real gap, such as a use case or competitor you did not cover.

As a reference point for structure, Semrush's prompt audit template organizes 27 prompts across four layers: discoverability, clarity, authority and trust [2]. That is one vendor's template, not a target size.

Example: a B2B accounts payable product

This is a hypothetical product, accounts payable automation for mid-size finance teams. [Brand], [Competitor A] and [Competitor B] are placeholders.

GroupPrompt
CategoryWhat are the best accounts payable automation tools for a mid-size company?
CategoryWhich software automates invoice approval workflows for NetSuite users?
Comparison[Competitor A] vs [Competitor B]: which is better for multi-entity companies?
ComparisonWhat are the best alternatives to [Competitor A] for invoice processing?
ProblemHow do we stop duplicate invoice payments?
ProblemOur month-end close is slow because invoices sit in email. How do we fix that?
BrandWhat is [Brand] and who is it for?
BrandDoes [Brand] integrate with NetSuite and how does pricing work?

Notice what the set does. The first six prompts can never return your brand by name-matching, so any appearance is earned. The last two check facts: integrations, pricing and positioning that a model may state wrongly.

Store the set in a simple sheet so changes are traceable:

id,prompt,intent,segment,added,status
C01,"What are the best accounts payable automation tools for a mid-size company?",category,mid-market,2025-01-10,active
P02,"How do we stop duplicate invoice payments?",problem,mid-market,2025-01-10,active
B01,"What is [Brand] and who is it for?",brand,all,2025-01-10,active

When you read results, separate a mention from a linked citation. The two mean different things, as covered in AI citation vs mention.

Refreshing the set

Refreshing means changing the list on purpose, not rerunning it more often. Those are two separate decisions.

Run frequency. Ahrefs says tracked prompts can be refreshed monthly, weekly or daily [1]. Pick the slowest cadence that still lets you notice change, since each refresh costs checks and review time.

List changes. Review the set on a fixed schedule, such as quarterly:

  • Keep a core. Leave most prompts unchanged so trends stay comparable.
  • Add prompts for new products, integrations, segments or competitors, and for questions that start appearing in sales calls.
  • Retire prompts that no longer match how buyers talk or that describe a product you no longer sell. Mark them retired in the sheet instead of deleting them.
  • Log every change with a date, so a shift in your results can be separated from a change in the list.

If a category or problem prompt consistently returns no mention of you, treat it as a diagnostic. It may point to a page you have not written, or to one that does not state the answer plainly. It is not a rule that every missing prompt needs a new article. Content should still be written for people first, which is Google's stated guidance [3].

What to do next

  1. Write 3 to 5 candidate prompts per intent group from real sales and support language.
  2. Remove your brand name from every category and problem prompt.
  3. Put the set in a sheet with intent, segment and date columns.
  4. Run it, read every answer, and note where you are absent, wrong or described oddly.
  5. If you want a quick manual baseline first, follow the steps in how to check if AI chatbots know your business.