Prompt-Set Libraries By Category: SaaS, DTC, B2B Services, And Local

Nobody wants to write sixty prompts from a blank document. What usually happens instead is that someone writes twelve in an afternoon, half of them phrased the way the marketing team talks rather than the way buyers talk, and the resulting measurement quietly encodes the company's own self-image. Starter sets solve the blank page. They introduce a different risk, which is copying someone else's assumptions about your category without noticing.

So these are published with their construction rules attached. The patterns below are deliberately generic, with bracketed slots you fill in. Adapt them, then check the adaptation against the rules at the end.

How these sets were built, and their version date

Set version: v1.0. Published 2026-09-01.

These starter sets were constructed from the prompt-set design method on how to build a prompt set: stage coverage, buyer phrasing, no brand baiting, and a fixed count per stage so that no single buying moment dominates the measurement. They are structural patterns, not measured findings. No result from any model was used to select or tune them, and none of them has been run as a study.

They are also, explicitly, not exhaustive and not universally applicable. Four commercial shapes are covered here. Regulated categories, marketplaces, hardware, open-source projects, education, and public sector all have buying journeys these patterns will fit badly. Treat a starter set as roughly half of a usable set. The other half is category-specific and only you can write it.

Version and date every set

This is the rule that people skip and then regret three months later.

Every prompt-set file needs, at minimum:

Field	Example value	Why
set_id	saas-core	Distinguishes parallel sets
version	1.2	Increments on any change to prompt text or count
created	date	Anchors the original
last_modified	date	Anchors the current state
changelog	dated list of edits	Says exactly what moved and when
locale	e.g. en-GB	Phrasing and results vary by market

The reason is trend comparison. When you compare a run in March against a run in June, you are assuming the only thing that changed between them was the world. If the prompt text changed, that assumption is false, and you cannot tell how much of the movement is your edit and how much is real. A single reworded prompt out of fifteen can move a mention rate on its own.

The discipline that follows:

• Never edit a prompt in place on a set you are trending. Increment the version and start a new series.
• If you must add prompts, add them as a clearly separate block and report the original block's numbers separately so the historical series stays intact.
• Retire prompts rather than deleting them. Mark them retired with a date, keep the text, and note that the series ends there.
• Re-baseline deliberately. When a version bump makes a comparison invalid, say so on the report rather than showing an unbroken line.

More failure modes of this kind are catalogued in prompt-set mistakes.

The anti-brand-baiting rule

This rule belongs in every set below, and in every set you build. Stated plainly:

A prompt must be one a real buyer could plausibly type without knowing your product exists. If a prompt describes a combination of attributes only your product satisfies, it is not measuring visibility. It is measuring your ability to write a prompt.

Baited prompts are seductive because they produce a number that makes the person who commissioned the measurement happy, and because they are easy to write accidentally. The usual accidental forms:

• Naming your own differentiating feature as a required constraint
• Using your internal category name instead of the buyer's
• Describing an audience so narrowly that only your customer base fits
• Including your brand name in a prompt that is supposed to measure unprompted mention

The last one deserves a note. Brand-named prompts are legitimate, but they measure something different: what a model says about you when asked directly. Keep them in a separate block and never mix them into an unprompted mention rate. Combining the two inflates the headline figure and it will not survive scrutiny from anyone who reads the method.

SaaS: prompts by buying stage

Generic patterns. Fill the brackets.

Discovery

• "What software do [role] use to [core job]?"
• "How do teams handle [recurring operational problem] at [company size]?"
• "Best tools for [job to be done] in [year]"

Shortlisting

• "Best [category] software for [company size or segment]"
• "Top [category] tools for [industry]"
• "[Category] platforms that integrate with [widely used adjacent system]"

Constraint-led

• "[Category] software that supports [deployment or data-residency requirement]"
• "[Category] tools with [pricing model, e.g. usage-based pricing]"
• "[Category] software for teams without a dedicated [specialist role]"

Alternatives and comparison

• "Alternatives to [widely known category incumbent]"
• "[Incumbent A] vs [Incumbent B] for [use case]"
• "Cheaper alternatives to [widely known category incumbent]"

Verification

• "Is [brand] good for [use case]?"
• "What are the limitations of [brand]?"
• "Does [brand] work for [specific constraint]?"

The incumbent slots take a genuinely well-known name in your category, not your closest competitor and not yourself. The point is to see who the model surfaces around a fixed reference point, and picking a reference point that flatters you defeats it.

DTC and ecommerce: prompts by buying stage

Consumer buying compresses the stages and leans heavily on constraints, occasions, and price bands.

Discovery

• "What should I look for when buying [product type]?"
• "Is [product type] worth it for [use case]?"
• "[Product type] for someone who [situation or need]"

Shortlisting

• "Best [product type] under [price band]"
• "Best [product type] for [specific use case]"
• "Most durable [product type] for [environment or frequency of use]"

Constraint-led

• "[Product type] that is [material, certification, or dietary constraint]"
• "[Product type] available in [region] with [shipping or returns constraint]"
• "[Product type] for [body type, size range, skin type, or equivalent fit constraint]"

Alternatives and comparison

• "Alternatives to [widely known consumer brand in the category]"
• "[Brand A] vs [Brand B] [product type]"
• "Dupes for [widely known product]" (include only if that phrasing is genuinely common in your category)

Verification

• "Is [brand] good quality?"
• "What do reviews say about [brand] [product]?"
• "Is [brand] worth the price?"

Two cautions specific to consumer categories. Price-band prompts go stale faster than any other pattern, so record the band as part of the prompt text and expect to re-baseline when it stops matching the market. And review-oriented prompts are downstream of third-party review content, which means a movement there often reflects a change on a review platform rather than anything you did.

B2B services: prompts by buying stage

Services buying is dominated by trust, scope, and geography, and it produces far more "how do I choose" prompts than "what is the best" prompts.

Discovery

• "Do we need [service type] or can we do it in-house?"
• "What does a [service type] engagement usually involve?"
• "When should a company hire a [service type] firm?"

Shortlisting

• "Best [service type] agencies for [industry]"
• "[Service type] firms that work with [company stage or size]"
• "How to choose a [service type] partner"

Constraint-led

• "[Service type] agencies in [region or market]"
• "[Service type] firms with experience in [regulated environment or specialism]"
• "[Service type] providers for a [budget model, e.g. retainer vs project]"

Alternatives and comparison

• "Agency vs freelancer for [service type]"
• "Alternatives to hiring a [service type] agency"
• "[Service type] agency vs building an in-house team"

Verification

• "Is [firm] a good [service type] agency?"
• "What is [firm] known for?"
• "What should I ask a [service type] agency before signing?"

The build-versus-buy prompts matter more here than in any other category, because they are where the model frames the whole decision. If the framing consistently favours in-house, that is a finding about the category's published discourse, and it is more actionable than any mention count.

Local services: prompts and regional variants

Local is where phrasing and geography stop being variants and become the measurement itself.

Discovery

• "How much does [service] cost in [city]?"
• "Do I need a licensed [trade] for [task]?"
• "How do I find a good [trade] near me?"

Shortlisting

• "Best [trade] in [city]"
• "Top-rated [service] in [neighbourhood or district]"
• "[Trade] in [city] with good reviews"

Constraint-led

• "Emergency [trade] in [city] available [timeframe]"
• "[Trade] in [city] that handles [specific job type]"
• "[Service] in [city] with [certification, insurance, or accessibility requirement]"

Verification

• "Is [business name] a reputable [trade] in [city]?"
• "What do people say about [business name]?"

Local sets need three deliberate axes rather than one:

1. Geography. One prompt per served area, at the granularity buyers actually use. City, then neighbourhood or district where that is how people speak, then the "near me" phrasing which behaves differently again because it depends on inferred location.
2. Urgency. Emergency and same-day prompts often surface an entirely different set of businesses than planned-work prompts. Measuring only one hides half the picture.
3. Job type. Trades in particular get segmented by task, and the model's answer for "leaking tap" may share nothing with its answer for "full bathroom refit."

Run the geographic axis with locale and region settings held constant and recorded, because an answer that depends on inferred location is not comparable across runs made from different places.

How to adapt a set to your category without breaking it

Variants: add them as a structured axis, not as extra prompts

The instinct is to write more prompts. The better move is to define variant axes and apply them systematically, so you can attribute movement to the right thing.

Axis	What varies	How to record it	What it tells you
Region	Locale, market, spelling	locale field on the run, not in the prompt text where avoidable	Whether visibility is market-specific
Phrasing	Formal vs colloquial, question vs command	Variant ID sharing a parent prompt ID	How sensitive the result is to wording
Specificity	Broad category vs narrow use case	Separate prompt with its own ID	Where in the funnel you appear
Persona	Role or seniority stated in the prompt	Variant ID	Whether framing changes by audience

Rules for variants:

• Change one thing at a time. A variant that changes both the region and the phrasing tells you nothing about either.
• Give every variant a parent ID. saas-core-07 and saas-core-07-uk should be visibly related in the file so you can report them together or separately.
• Keep run counts identical across variants. Comparing three runs of one variant against one run of another produces a difference that is mostly sampling noise.
• For regional variants, translate rather than transliterate. A prompt translated word-for-word into another market's language often stops being a phrase anyone would actually type. Have someone who buys in that market write it.
• Cap the total. Every variant multiplies your run cost. A set of 15 prompts with 3 runs each across 3 models is already 135 calls. Decide the ceiling before you start expanding.

The adaptation checklist

Before you freeze your version 1.0, check every prompt against these:

• Would a buyer who has never heard of us type this?
• Does it use their category noun, not ours?
• Does it avoid naming our differentiating feature as a requirement?
• Is it in the unprompted block, or is it a brand-named prompt kept separately?
• Is it stage-tagged, so no stage is over-represented?
• Does it have a stable ID that will not be reused if the prompt is retired?
• Is the locale recorded?

Then freeze it, date it, and leave it alone. A mediocre set you run consistently for six months produces better intelligence than an excellent set you keep improving.

The maintained starter sets live in the prompt-set library, with the worksheets and run logs in resources and downloads.

In Plain English

A starter prompt set saves you the blank page but not the thinking. Take the generic patterns for your category, fill in the brackets with your buyers' language rather than your marketing language, keep brand-named prompts in their own block, add region and phrasing variants one change at a time, then version and date the file and stop editing it.

This helps anyone about to measure AI visibility for the first time, and anyone whose existing set was assembled ad hoc and cannot support a trend. Use these sets as a starting draft in the four shapes covered. Do not use them as-is, do not assume they cover your category's real buying questions, and do not treat any of them as a complete list.

Next step: pick the closest category above, adapt fifteen prompts, run the checklist over them, and save the file with a version number and today's date.

Get the prompt-set template pack and start from a versioned file instead of a blank one.