peaksly explains · 04

Method

How do you choose useful prompts?

A prompt set is the fixed, versioned set of questions used to measure AI visibility. A good set does not repeat the same sentence twenty times. It reflects different situations in which a person wants to discover, compare or buy something, and only rarely searches for the brand itself.

6-minute readUpdated 5 Sep 2026with a fictional example

Prompt set for Mühlbach Solar · 20 search prompts · language DE · market Germany · version 1.0Fictional teaching example, not a measurement

Orientation8 of 20 How big does a PV system for a family home need to be? Is a battery worth it at 6 kWp? What does a heat pump cost in an old building? What funding is available for heat pumps? How much electricity does a 10 kWp system deliver per year? Does a heat pump work with old radiators? What does a wallbox with PV surplus charging bring? How long does an inverter last?

Comparison7 of 20 Heat pump or gas hybrid? Which batteries last longest? Air-to-water or ground-source heat pump for a terraced house? Which PV modules are worthwhile with shading? Regional installer or nationwide provider? Which providers offer PV with a battery as a package? Who installs PV and a heat pump together?

Purchase intent3 of 20 PV system quote Upper Bavaria Request heat pump package price Solar installer near me

Brand2 of 20 Mühlbach Solar reviews Mühlbach Solar maintenance

Every search prompt has exactly one search intent and a justified place. All questions and figures are fictional.

Set of 20 · version 1.0

Orientation: 8 of 20
Comparison: 7 of 20
Purchase intent: 3 of 20
Brand: 2 of 20

Brand questions deliberately at 2 of 20: they test your own name, not findability.

All names, questions and figures are fictional. The board shows a way of sorting by search intent, not a measurement and not a promise of results.

In short

Twenty questions, four intents, one version.

Useful search prompts follow real decisions of the target group. They cover different search intents, stay naturally worded and are documented as a fixed, versioned set for repeated measurements.

Step 1 · Basic termFictional example: Mühlbach Solar

A prompt is more than a keyword.

A classic keyword consists of a few words. A prompt often describes a complete situation: goal, context, condition and the decision at hand. That is why the same industry can appear in very different answers.

“heat pump old building cost” and “What does a heat pump cost in an old building of 140 square metres if the gas heating stays?” deal with the same topic but do not call for the same answer.

Naturally worded, not optimised. A prompt in the set sounds like a person’s question, not like a search engine line. Only then does the set measure what people actually ask.

Keyword

heat pump old building cost 5 words, a word list

The same words, but no situation: no floor area, no existing heating, no decision.

Prompt · search intent orientation

“What doesgoal a heat pump cost in an old building of 140 square metrescontext if the gas heating stayscondition?”

Decision the answer should support: hybrid or full replacement.

Step 2 · IntentsFictional example: Mühlbach Solar

The mix decides what the result can say.

A question set should not contain only broad provider searches. Comparison questions surface other competitors, purchase questions set other conditions, and brand questions test something entirely different: whether your own name produces an echo.

Not just your own brand name. A set of twenty brand questions almost always yields twenty mentions and says nothing about whether anyone finds Mühlbach Solar without the name. That is why brand questions deliberately stay at 2 of 20.

Two sets, two statementsfictional

Set A 20 × brand “Mühlbach Solar reviews”, “… prices”, “… maintenance” and 17 more brand questions 20 of 20mentions, because every question contains the name: an echo, not findability

Set B mixed 8 · 7 · 3 · 2 orientation, comparison, purchase intent, brand 9 of 20mentions, of which 7 of 18 without the brand name in the prompt

Brand deliberately at 2 of 20

18 of 20 questions without the brand name
mention readable separately per search intent
brand questions test presence, not findability

Step 3 · SelectionFictional example: Mühlbach Solar

From customer conversation to a verifiable set.

The questions come from customer conversations, enquiry forms and search suggestions. In the example, 34 questions were collected. Each receives a stamp: it stays in the set with its search intent, it is a variant of a question already present, or it describes an edge case hardly anyone asks that way.

In the end, 20 questions remain. Each has a justified place and can deliver its own finding.

No artificial variants. Swapping single words does not create a new search intent. Struck questions remain documented with their reason.

Review · excerpt 8 of 34 collected questionsFictional teaching example, not a measurement

1“What does a heat pump cost in an old building?”Customer conversationstays · orientation

2“Heat pump in an old building: what does it cost?”Search suggestionvariant of no. 1 · struck

3“Heat pump or gas hybrid?”Enquiry formstays · comparison

4“Is a battery worth it at 6 kWp?”Customer conversationstays · orientation

5“Does a 6.4 kWp system fit on a hipped roof with a dormer facing north-east?”Customer conversationedge case · struck

6“Solar installer near me”Search suggestionstays · purchase intent

7“Which batteries last longest?”Customer conversationstays · comparison

8“Which home battery lasts longest?”Search suggestionvariant of no. 7 · struck

8 of 8 reviewed5 stay3 struck

From conversation to set

Collected: 34questions
Variants struck: 9of 34
Edge cases struck: 5of 34
In the set: 20of 34

Step 4 · VersionFictional example: Mühlbach Solar

For developments, the set has to stay stable.

If the wording changes with every measurement, it remains unclear whether the answer or the question changed. Approved prompts therefore receive a fixed ID, language, market assignment and version. This set is what gets measured: in the Deep Audit as a snapshot, in monitoring continuously.

New questions start a new series. They are not retroactively counted into old results. Translations are checked as their own market and language version, not quietly treated as equal.

Version log of the setfictional

v1.0 20 prompts DE · market Germany · approved 14 Aug 2026, fictional Series 1runs 1 and 2

v1.1 20 + 2 prompts MS-21 and MS-22 added · from run 3 Series 2only for the 2 new ones, not retroactive

EN 1.0 20 prompts EN · market United Kingdom · own check of the version Series 3own market, not treated as equal

Prompt MS-07

Wording: “Who installs PV and a heat pump together?”
Search intent: comparison
Language, market: DE, Germany
Valid: since v1.0 · wording fixed

If the answer changes, it is not the question that changed.

Honest limitFictional example: Mühlbach Solar

What a good prompt set cannot do.

It never reflects every conceivable question in a market, and it does not guarantee a mention or a specific answer.

It does not replace a professional check of the answers and does not automatically make different AI search engines or measurement paths comparable. A good set is no oracle but a documented measuring instrument whose selection and limits remain traceable.

Not proven

every conceivable question in a market
a mention or a specific answer
professional correctness of the answers
comparability between AI search engines or measurement paths

What stays traceable

reason for selection per prompt
struck questions with their reason
version, language and market
wording fixed per series

What is measured with these prompts?

The fundamentals article separates mention, observed position and sources. The methodology shows the measurement paths.

AI visibility explained Open methodology →Look up the terms →