TL;DR: Your prompt set is the measurement instrument. Every AI visibility number you report describes the prompts you chose, so a badly built set produces precise measurements of the wrong thing. Build it from real buyer language rather than a keyword tool, cover four intents in proportion, test phrasing sensitivity, then version it and freeze it between comparisons.
Key Takeaways
Change the set, change the number. Adding prompts you win raises your rate with nothing improving. Version it, freeze it, and note the version on every chart.
Source from sales calls and support tickets. Prompts are conversational and longer than keywords, and keyword databases describe a different behaviour.
Cover four intents in proportion. Category discovery, comparative, problem-first and branded. Branded prompts flatter the number and belong in proportion, not in bulk.
Test phrasing sensitivity. Best X and affordable X are different queries. If you appear for one and vanish for the other, that is a finding worth acting on.
100 to 150 prompts for a single product in one market. Below about 50, variance dominates and monthly movement means nothing.
Review quarterly, change at boundaries. Language shifts and competitors appear. Change the set deliberately, at a quarter end, with the version recorded.
Anchor targets assume your CMS generates slug IDs from heading text. Confirm on staging and adjust the hrefs if your platform slugifies differently.
The set is the instrument
Every figure an AI visibility programme produces is a measurement of the prompts you chose. Not of your market, not of demand, not of what people actually ask. Of your list.
That makes the list the instrument, and instruments have to be built carefully and then left alone.
What goes wrong when it is not
Error | Consequence |
|---|---|
Prompts chosen because you win them | A flattering rate that moves with nothing |
Prompts chosen from a keyword tool | Measuring search behaviour, not prompting behaviour |
Mostly branded prompts | A high number that tells you nothing about acquisition |
Set edited mid-quarter | Your trend contains a methodology change nobody recorded |
One phrasing per question | Blind to the variant where you lose |
Set built once and never revisited | Measuring a market that has moved |
Each of these produces numbers that look fine. That is what makes them dangerous: nothing in a dashboard reveals that the instrument is pointed at the wrong thing.
The comparison worth making
If you ran a customer survey, you would care about the sample. Who was asked, how many, whether the questions were leading, and whether you changed the questionnaire between waves.
A prompt set is a questionnaire administered to a machine. Every one of those concerns applies, and almost nobody applies them.
What a well-built set gives you
A rate you can defend when someone senior asks how you got it. A trend that reflects performance rather than editing. And, most usefully, prompt-level findings: specific questions you lose, which convert directly into content briefs.
That last one is the real output. The aggregate rate is a summary; the individual prompts are the work.
The time it takes
A good first set takes about a day: a few hours gathering real language, a couple of hours structuring it, an hour testing phrasing variants.
That day determines the quality of every number you produce for the next year, which makes it the highest-leverage day in the whole programme.
Where prompts come from
From records of how people actually talk about your problem. Not from a keyword tool, which describes a different behaviour entirely.
The sources, ranked
Source | Why it is good | How to mine it |
|---|---|---|
Sales call recordings | The exact words buyers use before they are customers | Read transcripts of discovery calls; note the questions asked |
Support tickets and intake notes | Real problems in real phrasing | Pull the first line of tickets; that is the problem in the customer's words |
Lost deal notes | The comparisons that decided revenue | Note which alternatives came up and how they were framed |
Search Console queries | What already brings people to you | Longer queries especially |
Sales team interviews | Fast and surprisingly high quality | Ask three reps for the ten questions they hear most |
Community and forum threads | Unfiltered phrasing in your category | Read how people describe the problem, not the solution |
Why not a keyword tool
Keyword databases are built on typed search behaviour: short, fragmented, often two or three words. Prompts are conversational, longer, and frequently describe a situation rather than naming a category.
Someone searching types project management software. The same person asking an assistant says we are a six-person architecture practice and our current tool is too complicated, what should we use.
Those produce different answers, and only the second resembles how people actually prompt.
The intake question worth adding
Ask every new customer how they found you, and whether they used an AI assistant during their research. If yes, ask what they typed.
That single question produces better prompts than any other source, because it is your actual buyers reporting their actual behaviour. Most companies never ask it.
A practical first pass
Spend two hours reading ten sales call transcripts and fifty support tickets. Write down every question in the words it was asked. You will have forty to sixty candidate prompts, and they will be better than anything a tool would have given you.
Deduplicate, structure them by intent as the next section describes, and you have a first set.
The four intents
A representative set spans four intents, because your performance differs sharply across them and a set weighted toward one tells you about that one only.
The four
Intent | Shape | Share | What it measures |
|---|---|---|---|
Category discovery | best X for Y, what tools do Z | 30 to 40% | Whether you are in the consideration set |
Comparative | X vs Z, alternatives to Z | 20 to 30% | How you are positioned against rivals |
Problem-first | how do I solve Y, why does Y happen | 20 to 30% | Whether your content earns citations |
Branded | what is X, is X any good | 10 to 20% | How you are described |
Why the proportions matter
Branded prompts inflate everything. You will be mentioned in almost every answer to a question about your own name. A set that is half branded produces a high rate that moves with nothing and tells you nothing about whether new buyers find you.
Problem-first prompts are where content earns its place. These are the questions your articles answer, and the ones where a citation rate reflects your writing rather than your market position.
Comparative prompts are the closest to revenue. Someone asking X vs Z is deciding. Losing those matters more than losing a broad category question, and they deserve more runs per prompt.
Category discovery is the hardest and most valuable. This is the question a buyer who has never heard of you asks.
Weight by commercial value, not evenly
Within each intent, include more of what matters commercially. If three of your product lines generate most revenue, their prompts should outnumber the others, and the set should say so explicitly.
The test for a balanced set
Count your prompts by intent and compare to the proportions above. Then ask: if my rate rose five points, could I say which intent drove it?
If not, your set is not segmented enough to be actionable. Tag every prompt with its intent from the start, so every report can break down by it. That tagging takes ten minutes and it is what turns a single number into a diagnosis.
Phrasing sensitivity
Small wording changes produce genuinely different queries. Your brand can be the top recommendation for one phrasing and absent from the next, and almost nobody tests for it.
The test
Take one commercially important question and write six variants of it. Run each five times. Record where you appear and where you do not.
Variant type | Example |
|---|---|
Neutral | best project management software for architects |
Qualified by price | affordable project management software for architects |
Qualified by scale | project management software for small architecture practices |
Problem-framed | our architecture practice is struggling to track project deadlines |
Quality-framed | most powerful project management software for architects |
Simplicity-framed | simplest project management tool for a small architecture firm |
What the results tell you
Pattern | What it means | Action |
|---|---|---|
Present on neutral, absent on affordable | You are not positioned as good value anywhere models read | Publish honest pricing context, or accept the positioning |
Present on affordable, absent on powerful | You are bracketed as the budget option | A positioning finding for product marketing |
Present on all category variants, absent on problem-framed | You have no content answering the problem as described | Content brief, and the clearest one you will get |
Present everywhere | Strong position on this question | Move on |
Absent everywhere | Not in the consideration set at all | Entity and corroboration work, not content |
That third row is the most common and the most useful. Teams discover they rank for their category but vanish the moment a buyer describes their situation rather than naming the category, which is how most people actually prompt.
Build variants into the set
For your top ten commercially important questions, include two or three phrasings each rather than one. That costs thirty prompts and buys you the sensitivity view permanently.
For the rest, one phrasing is fine, chosen to match the most common real wording you found in your source material.
The qualifier words worth testing
Affordable, cheap, best, most powerful, simplest, easiest, enterprise, small business, for beginners, advanced. Each of these pulls an answer toward a different set of brands.
If a qualifier matters to your positioning, test it. If you consistently lose on a qualifier you believe describes you, that gap between self-perception and how the market is described is itself the finding.
What not to conclude
A single absence on one variant in one run is noise. Variants need the same multiple-run treatment as everything else, and a pattern across five runs is what counts.
How many prompts
Enough that variance does not dominate, few enough that you can act on the output.
Situation | Prompts |
|---|---|
Single product, one market | 100 to 150 |
Several products or segments | 200 to 400 |
Multiple brands or markets | Tier them, see below |
Local business, one location | 20 to 40 |
Testing the waters manually | 20 |
The lower bound
Below about 50 prompts, run-to-run variance dominates and month-to-month movement tells you nothing. You can still use a small set usefully, but report prompt-level findings rather than a rate, because the rate will bounce on noise.
The upper bound
Past a few hundred you are usually adding near-duplicates rather than coverage, and you are certainly adding cost. Tools price on prompt volume, and the triage work scales with the number of losses you have to read.
More prompts is not better. More runs per prompt is better, because it narrows your range and lowers the threshold at which you can detect real change.
Runs matter more than count
If your figures bounce too much to act on, add runs rather than prompts. Five runs per prompt is the practical minimum; ten for commercially critical questions.
Trade-off | Effect |
|---|---|
200 prompts, 3 runs | Broad coverage, noisy rates |
100 prompts, 5 runs | Balanced. The usual recommendation |
50 prompts, 10 runs | Narrow coverage, high confidence per prompt |
Tiering for multiple brands or markets
Do not multiply. Rank by revenue contribution and track accordingly.
Tier | Coverage |
|---|---|
Core brands and markets | Full set, full runs |
Secondary | Reduced set, same runs |
Monitored | 20 prompts, manual quarterly check |
That structure keeps the total manageable and gives you a defensible rationale when someone asks why a market is not fully covered.
A practical starting point
Begin with 100: roughly 35 category, 25 comparative, 25 problem-first, 15 branded, with phrasing variants on your top ten. Five runs each.
That is 500 runs monthly, manageable with tooling and at the edge of manual. Adjust after the first quarter based on what proved useful.
Versioning and freezing
The discipline that makes everything else meaningful, and the one most teams skip.
The rule
Give the set a version number and a date. Do not add, remove or reword prompts between measurements.
If you change it, your rate changes for reasons unrelated to performance, and the comparison across that boundary is meaningless. That is true even when the change is an improvement.
Why it is tempting to break
You will find a prompt you should have included. A competitor will launch. Someone will suggest a better phrasing. All reasonable, and all of them corrupt the comparison if applied mid-quarter.
Keep a parking list of proposed changes and apply them at the next boundary. That way nothing is lost and the trend stays valid.
What to record with the version
Field | Why |
|---|---|
Version number | v1.0, v1.1, v2.0 |
Date created | When this instrument came into use |
Prompt count, total and by intent | The shape of the set |
What changed from the previous version | So a step in the chart has an explanation |
Who approved it | Accountability for the instrument |
Marking version changes on charts
When you change the set, put a visible marker on the trend line at that point and note the version.
A step in your rate at a version boundary is then explicable rather than mysterious, and anyone reading the chart later knows not to compare across it. This single practice prevents a whole category of false conclusions.
When to create a new major version
Change | Version |
|---|---|
Adding a few prompts in an existing intent | Minor, v1.1 |
Rewording for accuracy | Minor, and note it |
Adding a new product line or market | Major, v2.0 |
Restructuring the intent mix | Major |
Changing more than about 20% of prompts | Major, and treat as a new baseline |
The honest consequence
A major version change means starting a new baseline. Your old trend remains valid for its period and the new one starts fresh.
That is a real cost, and it is the reason to spend the day getting the first set right rather than iterating your way to a good one while measuring.
Persona and context variants
Assistants adjust answers to the context they are given. Adding that context deliberately turns one prompt into several useful ones.
The variants worth adding
Variant | Example addition |
|---|---|
Company size | for a 10-person team, for an enterprise with 5,000 staff |
Budget | on a tight budget, we have budget for a premium tool |
Technical skill | we have no engineers, we have a strong technical team |
Industry | for a law firm, for a manufacturer |
Stage | we are just starting out, we are migrating from a legacy system |
Location | in the UK, in Germany |
Why these matter commercially
If you sell to small teams, prompts without a size qualifier measure a market you do not serve. Your rate on best project management software is interesting; your rate on project management software for a five-person team is the one that predicts revenue.
Adding the qualifier narrows the competitive field and usually reveals a very different picture.
The honest limit
You can generate hundreds of variants combinatorially, and most of them are not worth tracking.
Add persona context where it genuinely changes your competitive set. For most companies that means two or three dimensions, not six. If your answer does not change between for a 10-person team and for a 50-person team, do not track both.
Location, specifically
If you serve a geography, set location explicitly and keep it fixed. Assistants infer location and rewrite near-me phrasing into a city name, so an unset location produces answers for wherever your tooling happens to appear to be.
For multi-market businesses, treat each market as its own tier rather than mixing locations in one set.
How to use the findings
Persona variants produce the most directly actionable content briefs in the whole exercise. If you are absent from every prompt mentioning a small team, you need a page that says explicitly what you offer small teams, with the specifics.
That is a concrete brief derived from measurement, which is the output a prompt set exists to produce.
Maintaining the set
A set built eighteen months ago is measuring a market that has moved. Maintenance is quarterly, deliberate, and done at boundaries.
The quarterly review
Check | Question |
|---|---|
Fresh source language | Have new questions appeared in sales calls and tickets? |
Competitor set | Has a new entrant appeared in answers we do not track? |
Product changes | Have we launched, renamed or retired something? |
Dead prompts | Are any prompts producing nothing useful, every time? |
Intent balance | Has the mix drifted from the proportions we set? |
Phrasing | Has the way people describe this changed? |
An hour, once a quarter, with the parking list of proposed changes from the previous three months.
What triggers an off-cycle change
Very little should. The exceptions:
Trigger | Response |
|---|---|
You launch a new product line | Add prompts, new major version |
You rebrand | Update branded prompts, new major version |
You enter a new market | New tier, new version |
A prompt turns out to be ambiguous or broken | Fix it, note it, treat the boundary as a version change |
Competitor launches and interesting new phrasings go on the parking list. They are not emergencies.
Dead prompts
Some prompts never produce anything useful: too broad, too obscure, or phrased in a way no real person uses. After two quarters of nothing, retire them at the next boundary.
Retiring them is not the same as removing prompts you lose. A prompt you consistently lose is one of your most valuable, because it is a gap. Keep it.
The temptation to remove losses
Worth naming because it happens quietly. When a rate is under pressure, there is an incentive to drop the prompts dragging it down.
That is the measurement equivalent of marking your own homework. If you remove losing prompts, write it in the version notes and say so when reporting, or do not do it.
Keeping the record
One document per version, with the full prompt list, intents, tags, date and change notes. Store it somewhere that is not a vendor's platform.
Your prompt set is an asset you built from real customer language. It is portable, it survives tool changes, and it would take another day to rebuild.
A worked example
A fictional company: project management software for architecture practices, sold to firms of five to fifty people.
The set, v1.0, 104 prompts
Category discovery, 36 prompts
Example | Variants included |
|---|---|
best project management software for architects | plus affordable, plus simplest |
what software do architecture firms use for project tracking | one |
project management tools for small architecture practices | plus for a 10-person practice |
software for tracking architectural project deadlines | one |
Comparative, 26 prompts
Example | Note |
|---|---|
[Our product] vs [Competitor A] | Also reversed order |
alternatives to [Competitor A] | The incumbent |
alternatives to [Competitor B] | The budget option |
best [Competitor A] alternative for small firms | Persona variant |
Problem-first, 27 prompts
Example | Source |
|---|---|
how do architecture firms track billable hours across projects | Support tickets |
our drawings keep getting out of sync between team members | Sales calls |
how to manage RIBA stages in project software | Sales calls |
what happens when a client changes scope mid-project | Discovery calls |
Branded, 15 prompts
Example |
|---|
what is [Our product] |
is [Our product] good for small practices |
how much does [Our product] cost |
what are the downsides of [Our product] |
What this set produced
Illustrative, to show the shape of a useful finding:
Intent | Citation rate | Reading |
|---|---|---|
Branded | 94% | Expected. Tells us little |
Problem-first | 31% | Content is working on some topics, absent on others |
Category | 18% | Present but not prominent |
Comparative | 9% | The gap. Losing the decision-stage questions |
The comparative rate is the finding. The company has content answering problems and is losing the questions asked immediately before a purchase decision, which points directly at comparison pages.
The sensitivity finding
On the three variants of the main category prompt, the company appeared on neutral and simplest and never on affordable.
That is a positioning fact worth taking to product marketing, and it would be invisible in a set with one phrasing per question.
What it cost
A day to build, 520 runs a month at five runs each, and about ninety minutes of monthly triage. The comparative finding alone justified it.
Six prompt set failures
1. Built from a keyword tool. Keyword databases describe typed search: short, fragmented, two or three words. Prompts are conversational and often describe a situation rather than naming a category. Source from sales calls and tickets instead.
2. Too many branded prompts. You will be mentioned in almost every answer about your own name. A set that is half branded produces a high, stable, useless number. Keep branded to 10 to 20%.
3. Edited mid-quarter. Adding prompts you win raises your rate with nothing improving. Keep a parking list and apply changes at boundaries, with the version recorded on the chart.
4. One phrasing per question. You can be the top recommendation on one wording and absent on the next. Without variants on your important questions, you cannot see it.
5. No intent tags. Without them, a five-point move is unattributable. Tag every prompt at creation; it takes ten minutes and makes every report diagnostic.
6. Quietly removing prompts you lose. The prompts you lose are the most valuable ones, because each is a specific gap. Removing them improves the number and destroys the point of measuring.
The pattern
Four of six make your numbers look better than reality. That direction is not accidental: the pressure on anyone reporting a visibility metric is upward, and the prompt set is the easiest place to relieve it.
Which is exactly why versioning and change notes matter. They make the instrument auditable, including by you in six months.
The one to fix first
Sourcing. A set built from real buyer language is better than a larger, better-structured set built from keyword research, because it is measuring the right behaviour.
Two hours with sales call transcripts beats a day with a keyword tool, and the prompts you find will be ones you could not have invented.
FAQs
What is a prompt set?
The fixed list of questions you run against AI assistants to measure visibility. It is the measurement instrument: every rate you report describes the prompts you chose, so how it is built determines whether your numbers mean anything.
How many prompts do I need?
100 to 150 for a single product in one market, 200 to 400 across several products or segments, 20 to 40 for a single-location local business. Below about 50, variance dominates and monthly movement is meaningless.
Where should prompts come from?
Sales call transcripts, support tickets, lost deal notes and Search Console queries. Not a keyword tool, which describes typed search behaviour rather than conversational prompting. Two hours reading real transcripts produces better prompts than a day of keyword research.
What mix of question types should I use?
Roughly 30 to 40% category discovery, 20 to 30% comparative, 20 to 30% problem-first and 10 to 20% branded. Branded prompts inflate your rate and tell you little, so keep them proportionate and tag everything by intent.
Why does phrasing matter so much?
Because small wording changes produce genuinely different queries. You can be the top recommendation for best X and absent for affordable X. Include two or three phrasings for your top ten commercially important questions to see it.
Can I change the prompt set?
Yes, at quarter boundaries, with a version number and change notes, and a marker on the trend chart. Changing it mid-quarter means your rate moves for reasons unrelated to performance and the comparison is invalid.
Should I remove prompts I always lose?
No. Those are the most valuable prompts in the set, because each one is a specific, named gap you can act on. Removing them improves the number and defeats the purpose of measuring.
How often should I review the set?
Quarterly, for an hour, using a parking list of changes proposed during the previous three months. Check for new source language, new competitors, product changes, dead prompts and intent balance drift.
Building yours
Block out a day. Two hours reading sales call transcripts and support tickets, writing down every question in the words it was asked. Two hours structuring them across the four intents. An hour adding phrasing variants to your top ten. An hour tagging, versioning and writing it up.
That day produces the instrument every number you report for the next year will be measured with.
The checklist
Sourced from real buyer language, not a keyword tool
100 to 150 prompts for a single product in one market
Roughly 35% category, 25% comparative, 25% problem-first, 15% branded
Every prompt tagged with its intent
Two or three phrasing variants on the top ten commercially important questions
Persona context added where it changes your competitive set
Location set explicitly if you serve a geography
Version number, date and change notes recorded
Stored somewhere that is not a vendor's platform
Quarterly review scheduled, with a parking list for proposed changes
Then run it properly
Five runs per prompt minimum, ten for commercially critical ones, spread across a day, in fresh sessions, within the same window each month. Record the range alongside every rate.
If the output is a content backlog
A well-built set produces exactly that: a list of specific questions you lose, each one a brief. Getting those pages written is where most teams then stall, and Blazly SEO covers research, planning, drafting and direct publishing from about $39 a month.
The prompt set stays yours either way. It is built from your customers' language and it is portable across any tool you use.
Summary
Your prompt set is the measurement instrument, and every AI visibility figure you produce describes the prompts you chose rather than your market. Build it from sales calls and support tickets rather than a keyword tool, because prompts are conversational and keyword databases describe typed search. Cover four intents in proportion and tag every prompt so your reports can break down by them. Test phrasing variants on your most important questions, since appearing for one wording and vanishing for another is a finding you cannot see otherwise. Then version it, freeze it between measurements, and change it only at quarter boundaries with the change noted on the chart. The prompts you consistently lose are the most valuable in the set, so never remove them to improve the number.