How Do I Build a Prompt Library for Buyer Journeys?

A prompt library for AI visibility is a sample of the questions your buyers put to ChatGPT, Perplexity, Gemini and Google's AI answers, chosen so the numbers you get back describe your actual market. You build it from a handful of business questions, spread those across the stages of the buyer journey, vary a few dimensions on purpose, and give every prompt a role that decides how it counts in the report. Then you mostly leave it alone, which turns out to be the hard part.

Aleyda Solís made the case for this in her guide to a representative AI search prompt library: a library should be a structured sample of the journeys that matter to the business, not a tool's default list, and representative does not mean exhaustive. The seven journey stages below follow her guide. Other guides add useful pieces, such as the layer weights in ai-visibility-software.com's prompt-set method and the ownership and status tags in HubSpot's AEO prompt tracking guide.

What I couldn't find in any of them was a complete worked set, labelled, with the rules for what each label does to the numbers. So that's what this post is. The example category is ours, AI visibility tools for B2B SaaS, because I know how its buyers talk. Swap in your own.

Start from business questions

A prompt nobody can connect to a decision is noise you pay to collect. So before writing a single prompt, write down the four to six questions the business actually wants answered. For our category they'd look like this:

1. When a content lead at a small B2B SaaS company shops for an AI visibility tool, do the engines name us?
2. When a buyer weighs us against Profound, Peec AI or Otterly, what do the engines say?
3. Do the engines describe our pricing, plans and engine coverage correctly?
4. When someone is still working out the problem ("why does ChatGPT recommend my competitor"), are we part of the answer?
5. After someone signs up, do the engines help them use the product, or send them somewhere else?

Every prompt in the library points back at one of these. When a prompt can't, it goes.

The seven journey stages

Each stage asks a different kind of question, and not every stage produces answers that name vendors. That matters, because our own AI rank tracker guidance comes down to one test: if the answer to a prompt would name a vendor, track it; if the answer would be a definition, it's a blog topic. Discovery prompts fail that test more often than any other stage, so I keep only a few.

Evaluation and comparison carry the most weight in a buying category because those answers decide the shortlist. In the worked set below they hold 8 of the 20 core prompts, plus all 8 competitor prompts.

The dimensions you vary on purpose

A dimension is anything a buyer adds to a question that could change which brands come back. The nine I use:

The rule that keeps this manageable: vary one dimension per prompt, two at most. A prompt with four modifiers ("cheapest self-serve AI visibility tool for a German fintech that connects to Search Console") gets a narrow answer, and when you lose on it you can't tell which modifier did it.

One business question, several prompts, no near-duplicates

Take business question 1: do the engines name us when a content lead at a small B2B SaaS company shops for a tool? Written once, it becomes one prompt and one number. Written well, it becomes five prompts, each testing something the others can't:

And here are two I'd cut:

• "top AI visibility software for SaaS companies"
• "which AI visibility platform is best for B2B software"

Both are the first prompt with different words. The test I use has two parts: would you expect a different set of brands in the answer, and would a loss make you do something different? If both answers are no, it's a duplicate. Duplicates are worse than useless, because they double-count whatever happened on the original and make one question look like three.

Synonym variants do have a use, as experiments. If you suspect "platform" and "tool" pull different answers in your category, put one variant in the experimental pool for a month and look. Most of the time the answers match and the variant gets retired.

Five roles, and what each does to the numbers

This is the part I think most prompt lists skip. Each prompt gets exactly one role, and the role decides where its result shows up in a report.

Competitor prompts are split out because the rival's name pulls the answer toward that rival, so a win there means something different from a win on a core prompt.

Branded prompts would wreck a headline rate if mixed in. On our own account in August 2026 the mention rate was 6% on branded prompts and 0.7% on unbranded ones. CueScout keeps branded prompts out of its totals by default for that reason, and I'd do the same in a spreadsheet. Their real job is business question 3: does the engine get your pricing, plans and coverage right?

Monitoring prompts are the odd ones. A definition question where you want to see whose framing the engine borrows, or a "reviews of [your brand]" prompt where you're checking for wrong facts. Keep them few, because somebody has to read each answer.

The worked set: 40 prompts

Twenty core, eight competitor, four branded, three monitoring, five experimental. Fill in your own category; the shape is the point.

A few choices worth explaining. C19 is the one product-line prompt in core, because a buyer tracking several brands prices the tool differently from a team with one; if you sell one thing to one kind of buyer, drop it. B03 is there because engine coverage changes, and I want to know when an engine is still repeating an old version of it. X05 deliberately breaks the two-modifier rule, as a test of whether urgency changes anything at all. My guess is it doesn't, and in a month the data will say.

The range is 30 to 50 because both ends have a failure mode. Below 30, the comparison stage or a single competitor rests on one prompt, and one odd answer moves the whole line. Above 50, for one brand in one market, the extra prompts are usually near-duplicates nobody noticed, and nobody reads the review.

Tag every prompt the same way

Whether the library lives in a spreadsheet or a tool, every prompt carries the same fields. This is the record I keep for C11:

id: C11
prompt: AI visibility tool that connects to Google Search Console
role: core
stage: evaluation
varies: integration=search-console
business_question: BQ1
source: sales-call
added: 2026-10-13
owner: content
status: active   # active | paused | retired

The source field gets ignored more than any other, and it's the one that tells you at the next review whether a prompt came from a buyer or from someone's head. Sales-call notes, support tickets, the questions buyers repeat in public threads, competitor comparison pages, and Search Console all count as sources. "We thought of it" is a source too, as long as you're honest that it is.

Where Search Console fits

Search Console is useful here, with limits that are easy to forget.

It lists Google queries your site got impressions for. That is wording evidence: it shows the modifiers people attach ("for small teams", "free", "vs") and the vocabulary they use instead of yours. Borrow from it freely.

What it can't do is measure what anyone asks an assistant. It only contains queries where you already appear on Google, so it's filtered by your existing rankings, and Google searches are shorter and differently shaped than the questions people type into ChatGPT. No public source measures assistant query volume, so don't weight prompts by impressions as if it did.

And clean it before you read it. On cuescout.com, 87 of 777 impressions over three months came from queries stuffed with -site: operators, which is scrapers at work. I wrote up how to spot competitor-monitoring bots in Search Console. A library built on unfiltered data ends up tracking questions no buyer typed.

If you want more on the sources side, the AI prompt research page lays out how we combine suggested prompts, repeating public questions and Search Console into one starting list.

The refresh protocol

The library is only worth something if this month's numbers can be compared with last quarter's. That needs a stable core and a controlled way in for new prompts.

1. Never edit core prompt text. A reworded prompt is a different prompt with no history. Add the new wording as experimental and pause the old one if it gets promoted. CueScout enforces this: prompt text can't be edited, so a rewording starts fresh and the old prompt keeps its history while paused.
2. Every month, decide on each experimental prompt: promote, extend for another month, or retire. Promote only if it maps to a business question and doesn't duplicate an existing core prompt.
3. Every quarter, review the core. Retire a core prompt only when its business question is gone, such as a dropped product line or a market you left. Losing on a prompt is not a reason. That loss is exactly the reading you're paying for.
4. Cap the churn. I'd swap no more than a handful of core prompts in a quarter. If you're changing more than that, the business questions changed, so rewrite those first.
5. Route events through the experimental pool. A launch, a new competitor in your deals, a pricing change: each gets experimental prompts first, and they earn their way into core at the review.
6. Log every change with a date and a reason. When you report a trend, compare only the prompts present in both periods, and say how many that was.

That last rule is the one that saves you in front of a client or a CEO. "Visibility went up four points" means nothing if a third of the prompts changed in between.

Fitting 30 to 50 prompts into a plan

Every tool caps prompts somewhere. This is how the set fits ours.

Starter has 20 prompts, checked daily on ChatGPT, Perplexity and Google AI Overviews. That holds one brand's core set and nothing else, so the 40-prompt library doesn't fit, and I won't pretend it does. On Starter I'd track the 20 core prompts, keep the rest in the spreadsheet, and run the branded and monitoring prompts by hand once a month. Starter allows two brands, but each brand reserves at least 10 prompts, and 10 is too thin for a journey library. One brand is the realistic use.

Growth has 100 prompts and adds Gemini and Google AI Mode. The full 40 fits with 60 to spare. I'd spend the spare on one more market or language as its own core set, or on a second brand, before I'd let the experimental pool grow past ten. More prompts in one market mostly means more near-duplicates.

Two add-ons change the engine list without changing the prompt count. Claude checks every tracked prompt daily. The Grok pack asks Grok 20 prompts you pick, weekly, and I'd pick them from core so Grok's numbers line up with the rest.

Inside CueScout, the roles go in as prompt tags, and the intent label (buying, comparison, informational, branded) is filed automatically by rules. Branded prompts already stay out of the headline rates until you add them back.

What to do with the results

A library that only produces a rate is half done. Where the core prompts show you missing, read the pages the engines cited instead, and turn each gap into a piece of work: a page to fix, a question to answer on your own site, a list to get onto. That step is what planning content from AI answer gaps covers, and if you haven't collected cited pages yet, how to audit brand mentions in AI engines is the first pass.

One caveat I'd keep in mind throughout. A prompt library is a sample of questions you chose, and every number it produces is about those questions. It can't tell you how many people ask them. Choose them carefully and the sample is useful, and it still doesn't measure demand.