How to Build a Prompt Set That Actually Measures AI Visibility

Adding and deleting prompts every sprint makes week-eight incomparable to week-one.

The trend line is the product.

A rotating set is a series of one-off anecdotes.

Freeze the core.

When you must change it, version the file with a date (prompt-set-v2026-08-21.csv) and note which IDs changed.

Too small — can n=10 tell signal from noise?

The same prompt, same engine, minutes apart, can name a different vendor.

That is how the models work.

With ten prompts you cannot tell a real loss from a redraw.

Thirty to fifty core prompts is the range that stays cheap enough to re-run and large enough to trend.

Two hundred is mostly cost.

Confirm you are actually invisible before you industrialize the set — the 7-layer diagnostic is the pre-check.

What is the three-stage distribution model?

LLMConquer tags prompts tofu, mofu, and bofu.

Those map to Awareness, Consideration, and Decision — the same topic names used in onboarding.

Onboarding in the product defaults to 40 / 40 / 20 (Awareness / Consideration / Decision) because a first workspace needs category coverage, not only bottom-funnel queries.

You can change those weights.

For a commercial core index I still recommend Decision-weighted, because that is the leading indicator a revenue team can act on.

Brand names belong on Decision prompts only when you are deliberately testing a branded index.

In onboarding, generated prompts follow the same rule: no brand name unless the stage is BoFu.

Where should you source prompts?

Six methods, ranked by yield.

1. Sales call recordings

Highest yield.

Buyers already asked the question in their words.

Transcribe the last 20 discovery calls.

Extract questions, not your answers.

One intent per line.

2. Competitor gap

Prompts where a competitor is named and you are not.

This is the fastest way to stop measuring vanity.

In LLMConquer this is the competitor mode in AI Research (Pro).

3. Support tickets and on-site search

People who already know you still ask comparison and "does it do X" questions.

Those often belong in Consideration, not Awareness.

4. Existing keyword data, rewritten

Useful as a backlog, harmful as a paste.

Rewrite rules:

• Turn the fragment into a grammatical question
• Add the job to be done ("for an agency", "for five engines")
• Remove your brand from the core index
• Drop anything you would never say out loud

5. AI-generated variations of prompts you already trust

Once you have 15 good seeds, generate paraphrases and then semantically dedupe.

String-unique is not unique.

This is the similar mode in AI Research.

6. Category surface

Ask the engine what it volunteers about the industry with no vendor list in the prompt.

Capture the questions it implies.

This is the niche mode in AI Research.

What are the hygiene rules?

• One intent per prompt.
• No brand name in the core index. Keep a separate branded index.
• Natural phrasing, not keyword syntax.
• Locale and language set explicitly (en-US, en-GB, fr-FR). Do not assume.
• Deduplicate by meaning, not by string.
• Freeze the core set. Version any change with a date.

How many prompts do you actually need?

Variance is the constraint.

One answer is a draw from a distribution, not a census.

n=10 cannot separate a 8-point real drop from a bad afternoon.

n=30–50, re-run daily or weekly, with a 7-day rolling mention rate, is enough for most teams to see direction.

n=200 multiplies engine cost and refresh time.

You learn more from a frozen 40 than from an ever-growing 200 you cannot afford to re-run.

Starter plans in this category often cap prompts (LLMConquer Starter includes 25).

If you are under a cap, cut Awareness first, not Decision.

How often should you re-run, and what is a real change?

Daily is enough for a 30–50 set on the engines you sell against.

Hourly sampling changes the number without changing the decision.

A real change is a shift that survives a rolling window and shows up in relative share, not only in one engine's one-day mention %.

Absolute percentages disagree across tools.

That is expected.

Compare trend and rank order.

Do not fire a vendor — or a teammate — over a single-day dip.

Why two dashboards diverge is in Why two AI visibility tools give you different numbers.

A starter template you can use today

Forty prompts.

Stage-tagged.

Category-generic so a B2B software team can paste and then replace the Decision lines with their own market language.

Download the CSV — no email gate.

Replace the Decision rows with your category ("HRIS", "warehouse WMS", "payment orchestration") before you treat this as your index.

Keep the shape.

If you have not confirmed a real gap yet, start with the 7-layer diagnostic.

Best AI Visibility Tools in 2026: 12 Platforms Compared on Real Cost Per Tracked Prompt

Twelve AI visibility platforms compared on a normalized cost: monthly price divided by included prompts × engines × refreshes per week. Prices checked 21 Aug 2026.