How to build a prompt set for AI visibility

What does a real prompt set look like?

Here is ours, warts included. Attensira tracks 61 questions for attensira.com and reads them on ChatGPT and Perplexity on a daily schedule; in the 31 days from 4 September to 4 October 2026, answers were stored on 25 of those days. This chapter's window runs one day longer than the 4 September to 3 October window on our other pages because it includes 4 October's answers, so its counts are slightly higher (197 of 198 branded answers here, 180 of 180 there). This is how the questions were grouped on 4 October 2026, read over the Attensira MCP, with the answers each group produced in that window.

• Product category searches

  Prompts

  7

  Answers read

  769

  Example question

  best GEO tools

• GEO how-to and education

  Prompts

  6

  Answers read

  651

  Example question

  how to get cited by ChatGPT

• Audience and use case

  Prompts

  6

  Answers read

  629

  Example question

  AI search visibility for agencies

• Comparison and alternatives

  Prompts

  4

  Answers read

  445

  Example question

  Peec AI alternatives

• Not showing up in AI

  Prompts

  3

  Answers read

  334

  Example question

  why is my brand not mentioned by AI

• Pricing and budget

  Prompts

  2

  Answers read

  211

  Example question

  how much does AI search visibility cost

• AI crawlers and bots

  Prompts

  1

  Answers read

  105

  Example question

  AI crawler logs analysis tool

• Brand and reputation

  Prompts

  0

  Answers read

  no prompts filed

  Example question

  none filed here

That table holds 29 of our 61 prompts. The other 32 were added later under labels that never made it onto the topic board, and our three branded questions, "what is Attensira", "Attensira reviews" and "Attensira vs Profound", are among them rather than under Brand and reputation, which is why that row reads "no prompts filed" rather than a number. "Not measured" is kept for an engine nobody read.

We show the mess on purpose, because each flaw in it is one of the mistakes this chapter is about, and we made all of them.

The branded questions were mixed into the total. They were named in 66 of 66, 66 of 66 and 65 of 66 answers: all but one. Every unbranded question was named in none. Our workspace-wide share of voice is therefore entirely the branded group.

The set grew at different times. Prompts we tracked from early September had 105 to 124 answers each in the window. Prompts added in mid-September had 60 to 66. Prompts added at the end of September had 18 to 30. Attensira does not backfill a new prompt, which is correct, but it means a rate on a new prompt rests on a fraction of the evidence behind an old one.

The grouping drifted. Half the set sits outside the topic board, so a topic-level number describes the half that was filed, not the set.

None of this makes the readings wrong. It makes some of the obvious summaries of them wrong. The rest of this chapter is how to avoid building a set like ours on day one.

Why is the prompt set the most important measurement decision?

Because every number you report is an average over it. Share of voice, citation rate, position and sentiment are all computed over the answers to a specific list of questions. Change the list and you change the number, even if nothing about the engines changed.

In survey terms, the prompt set is your sampling frame. A frame that leans toward easy questions produces a flattering number; a frame that leans toward hard ones produces a gloomy one. Neither is a measurement of how buyers are advised unless the frame looks like the questions buyers actually ask. Two companies with identical presence in the engines can report very different visibility because they chose different questions.

So the set deserves more care than any dashboard setting: write it down before looking at results, justify each question by where it came from, and treat changes to it as events with dates.

What makes a good prompt?

A good prompt is a question a real buyer would type into an assistant, in their words, before they know your name. Attensira's own guidance puts it plainly: write prompts in the words your buyers use, not your own category language, and phrase each one as a complete question.

Compare three versions of the same intent:

• Keyword fragment

  Example

  expense tool agency

  What it measures

  Little; assistants give thin, unnatural answers to fragments

• Category language

  Example

  best accounts receivable automation platform

  What it measures

  How the engine answers people who already speak your vendor's vocabulary

• Buyer question

  Example

  what's the best way for a 20-person agency to stop chasing unpaid invoices?

  What it measures

  How a buyer with the problem is actually advised

The third is longer, specific to a situation, and free of jargon. It is also the question that produces an answer naming the products a buyer would be steered toward. Situational detail matters: "for a 20-person agency", "without a finance hire", "that works with Xero". Buyers include those details, and engines change their recommendations because of them.

Leave your brand name out of the prompt. Attensira's documentation calls an unbranded prompt the honest test of whether a model reaches for you unprompted. Branded questions are worth tracking too, but as their own group, for a different purpose: checking that the engines describe you correctly.

Where should the questions come from?

From evidence, not imagination. Good sources, in roughly descending order of value:

• Sales calls and demos. The question the prospect asked before they found you is the prompt.
• Support tickets and onboarding questions. These become retention-stage prompts: how to connect, fix or leave a product.
• Search Console queries that are phrased as questions, especially long ones. They show the words people already use.
• Competitors' pages that answers cite. If a comparison article decides an answer, the question it answers belongs in your set.
• Community threads where buyers ask for recommendations in your category.
• Pages AI crawlers fetch from your site that no tracked prompt explains. Attensira's weekly discovery uses this signal and competitors' cited URLs to propose at most three new questions a week, and adds none without your approval.

Write each question down with its source. When someone asks later why a question is in the set, the answer should be a fact, not a feeling.

How should a prompt set cover the buying journey?

A set made only of "best X" questions measures the shortlist and nothing else. Buyers ask different questions at different stages, and engines answer them from different sources.

Attensira files every prompt by intent, what the person wants from the answer, and stage, where they are in the buying journey:

• Awareness

  Typical intent

  informational

  Example

  how do agencies get paid faster?

  Why track it

  Shows whether you are part of how the problem is explained

• Consideration

  Typical intent

  commercial

  Example

  best invoicing tool for small agencies

  Why track it

  The shortlist question most teams track first

• Decision

  Typical intent

  comparison

  Example

  Ledgerly vs Xero for agencies

  Why track it

  Decides between named products, often from review sites

• Retention

  Typical intent

  transactional

  Example

  how to connect Ledgerly to Stripe

  Why track it

  Your own docs usually decide it, and errors here cost customers

A balanced set has questions at every stage, weighted toward the ones that decide purchases in your category. Reporting by stage then tells you something useful: being absent from awareness answers is a content problem, absent from consideration answers is a third-party presence problem, and wrong in retention answers is a documentation problem.

How many prompts do you need?

Fewer than you think, read more deeply than you think. Each prompt is read on every engine you track, at most once a day on its own slot, with one run per reading on Starter and three, issued at the same moment, on the higher plans. A prompt is a stream of observations, and a rate needs enough of them to mean something.

A practical way to size the set:

1. Start from the decisions you want to make. If you will act per topic, each topic needs enough prompts to pool a meaningful sample. Five prompts read daily on two engines for a month is about three hundred answers, enough to see large differences.
2. Prefer depth over breadth. Fifty questions sampled three times per reading will detect smaller changes than a hundred and fifty sampled once.
3. Leave headroom. Attensira sells prompts as slots held at once, not a monthly allowance; each plan's slot count is on the pricing page. A slot returns when you remove a prompt. Keep a few free so you can add the question a lost deal reveals without dropping one you need.

How do you change a prompt set without breaking your history?

Treat it like a change to a database schema: deliberate, dated and documented.

• Log every change with its date and reason. Our own ChatGPT share of voice jumped from zero to a steady positive reading the day our three branded prompts started producing answers. Nothing in the engines changed that day. Without the log, it would read as a win.
• Compare like with like. Across a change, compare only the prompts present on both sides. A total that includes new prompts compares two different samples.
• Keep branded and unbranded apart in every report, before and after.
• Do not edit a prompt's wording in place. A reworded question is a new question. Remove the old one and add the new one, so its history starts honestly.
• Retire questions on purpose. Remove a prompt when it stopped mattering, when two prompts turned out to be the same question, or when its answers never name anyone in your category.

What should you check before you commit the set?

• Every question is phrased as a buyer would type it, with no brand name unless it belongs to the branded group.
• Every question has a recorded source.
• The set covers awareness, consideration, decision and retention, weighted to how your buyers decide.
• Branded questions are a separate group.
• Every prompt is filed under a topic, and no topic is empty.
• The set is small enough to read deeply on every engine you track.
• There is a dated log for changes, starting today.

The prompt ideas tool and the query fan-out tool help turn a short list into buyer-shaped questions and the sub-questions engines ask behind them. Then read the next chapter, on a defensible report, before you show anyone a number built on the set.