Writing good prompts

Prompts are the input that determines whether your data is useful. A good prompt tracks how AI engines answer a question your customers actually ask. A bad prompt tracks how AI engines answer a question nobody asks, or how they parrot back your brand name when prompted with your brand name.

This guide covers what makes a prompt useful, what to avoid, and worked examples by industry.

The one rule

Write the question your customer types, not the question you wish they'd ask.

That's the whole guide in one sentence. The rest of this document is detail.

What a useful prompt looks like

A useful prompt has four traits:

It's a real question. Phrased the way someone would actually type it into ChatGPT or Perplexity. Not a search-engine keyword stub, not a marketing tagline.
It's category-level, not brand-level. "Best CRM for small B2B teams" is category-level. "Is HubSpot good?" is brand-level — it forces the engine to answer about HubSpot specifically, so it doesn't tell you whether your brand competes.
It has commercial intent. Someone asking it is closer to a purchase decision than to idle curiosity. "Cheapest way to send invoices" beats "What is invoicing software."
The engines actually answer it. Some prompts get refused (medical, legal, anything personally identifying). Some get vague "it depends" answers. Drop those.

What a bad prompt looks like

Three patterns to avoid:

Brand-locked prompts

What is Stripe?
Is Stripe a good payment processor?
Stripe pricing

These force the engine to talk about Stripe. They'll all return Stripe-positive answers. You'll learn nothing about whether the engine recommends Stripe over competitors when the user is shopping.

Vague informational prompts

What is fintech?
How do payments work?
Tell me about online business

The engine will answer with a paragraph of encyclopedia text. No brand mentions, no rank, no signal.

Branded long-tail SEO leftovers

Stripe vs PayPal vs Square for SaaS billing in 2026
Best Stripe alternative for high-risk merchants in Texas

These look detailed but they're contrived. Real users ask shorter, messier questions. The engine will either refuse the precision or answer with a flat list that doesn't reflect how anyone actually decides.

The decision-shaped prompt

The best prompts are shaped like decisions. The user is choosing between options and asking the engine for help. Examples:

"best CRM for a small B2B sales team"
"what's the cheapest way to send an invoice"
"alternatives to Salesforce for under 50 users"
"should I use Stripe or PayPal for a SaaS startup"
"tool to track AI search visibility"

Each of these gets a comparative answer. The engine names two to six options, often with brief pros and cons. That's the data you want — because that's the data that maps to a sales-funnel decision.

How many prompts you need

The minimum useful set is 3 well-chosen prompts. That's why Starter ships with 3 slots.

The honest answer for most companies: 8-15 prompts cover the bulk of your decision-shaped queries. After 15, you hit diminishing returns. After 30, you're tracking variations of the same question.

If you're on Agency with 50 slots, that's enough to cover multiple personas or geographies. Don't use the slots just because they exist — empty slots are fine.

Variations: when to track them

Two prompts that mean the same thing can return different answers across engines. Three useful variation patterns:

Synonyms

best CRM for small businesses
best sales software for small businesses
top CRM tools for SMBs

Run all three if your prompts are cheap (Growth, Agency). On Starter, pick one and move on.

Persona shifts

best CRM for a small B2B sales team
best CRM for a solo founder
best CRM for a 200-person enterprise sales org

These return different answers. If your product serves multiple personas, track at least one prompt per persona.

Geographic specificity

best payment processor for startups
best payment processor for startups in the UK
best payment processor for startups in Singapore

Useful if your geographic mix matters. Skip if you sell globally with no regional differentiation.

Industry examples

B2B SaaS (CRM example)

Decent set of 5 prompts for a CRM challenger:

best CRM for a small B2B sales team
Salesforce alternative for teams under 50
cheapest CRM with pipeline management
CRM with the best Gmail integration
should a startup use HubSpot or a simpler CRM

Each is decision-shaped. Each names a category. None forces a specific brand into the answer.

E-commerce (Shopify accessory)

For a Shopify app or theme:

best Shopify app for upsells
how to add a subscription product to Shopify
Shopify themes that load fastest
free Shopify apps worth installing in 2026
best Shopify alternative for digital products

The last one is interesting — it tracks how engines handle "alternatives to" queries, which often surface lesser-known brands.

Marketing agency

For a positioning-and-content agency:

best B2B marketing agency for SaaS startups
how to hire a content agency for a Series A company
agencies that do GEO (generative engine optimization)
fractional CMO vs marketing agency
marketing agency for AI startups

Note that some of these (the last two especially) are decision-shaped queries where the engine may name service categories more than specific firms. That's data too — it tells you the engine isn't yet treating your category as a named-brand space.

Local services (Ghana-based plumber example)

how to find a trusted plumber in Accra
emergency plumber near me Accra
average cost of plumbing repair in Ghana
how to verify a plumber's credentials
plumber for new building installation Tema

Local prompts return very different answers across engines — Perplexity tends to cite local directories, ChatGPT tends to generalize. Track at least two engines worth of variation.

Consumer apps (expense tracker)

best offline expense tracker app
expense tracker that works without internet
private expense tracker that doesn't sync to the cloud
Mint alternative for personal use
how to track expenses on Android without an account

The "without an account" / "without internet" angle is a useful prompt category for any privacy-positioned consumer tool — those questions get asked, and the engines answer with named brands more often than you'd expect.

When the engine refuses to answer

Some prompts trigger content-policy refusals on one or more engines. Common patterns:

Anything implying medical advice ("best supplement for…")
Anything implying legal advice ("which lawyer should I hire for…")
Anything political or naming a specific elected figure
Anything that pattern-matches as defamation ("worst CRM" with named competitors)

When a refusal happens, the snapshot finalizes with the other engines and marks the refused engine. If three of four engines refuse a prompt, drop it. The data won't stabilize.

When the prompt is too obscure

A prompt nobody is asking will return either a vague answer or a flat refusal across all four engines. After two weeks of empty snapshots, drop it.

This is most common with hyper-specific feature comparisons ("CRM that supports OAuth 2.1 with custom claim mappings"). Real users don't phrase questions like that. Engineers do — but engineers also don't ask ChatGPT to recommend tools at that level of specificity.

Editing prompts vs deleting them

If you edit a prompt, historical snapshots stay attached to the old version. The dashboard shows a small "edited" marker on the prompt timeline so you don't confuse pre-edit and post-edit data.

If you delete a prompt, snapshots are retained for 30 days then purged. Export anything you want to keep before deletion.

On Starter, deletes free up slots immediately. We don't enforce a cooldown.

A short checklist before saving a prompt

Does this sound like something a real customer would type?
Is it category-level, not brand-level?
Would the answer help someone make a decision?
Have I tried it once in ChatGPT manually to see if the answer is useful?
Does it fit my tier's slot count, or do I need to drop a weaker prompt first?

If you can answer yes to all five, save it. If not, rework it before it eats a slot.