The quality of your GEO measurement depends entirely on the quality of your prompts. A poorly written prompt produces an unreliable result — not because the AI system fails, but because the question doesn’t match how real users search.
Principle 1: Write like a user, not like a marketer
The biggest risk: writing from an internal perspective. A user doesn’t ask “What are the USPs of product X?” but “Which savings account currently has the best interest rate?” or “Which company inspects my fire extinguishers?”.
Principle 2: Vary the phrasing
The same question, phrased three different ways, gives three different results. Variation yields a more robust picture — but mind what you vary. These three cover different intents (maintenance, contract, comparison) and each belongs in your set:
- “Which company does annual inspection of fire extinguishers in Flanders?”
- “Where can I take out a maintenance contract for fire safety?”
- “Compare Belgian fire safety installers”
If you want to measure whether your presence hangs on phrasing (Fragility, see 4.3), you need something else: variants that ask exactly the same thing in other words — same category, same market, same decision question.
Principle 3: Cover all intent types
A good prompt set contains prompts for each intent type. Anyone who only tests comparison prompts misses the informational domain.
Principle 4: Use 'real intent' prompts
Also test raw, unpolished questions the way real users type them: “fire extinguisher inspection flanders” or “which company good for fire safety hospitality”.
Principle 5: Document and stabilise
A prompt set is only comparable over time if it is stable. Document each prompt with an ID and don’t change it unless there’s a good reason.
Before/after example: bad FAQ
Q: What do you do?
A: We offer professional fire safety solutions tailored to your needs with personal service. Feel free to contact us for a no-obligation quote.
Before/after example: good FAQ
Q: What maintenance contracts does [Company name] offer for fire extinguishers, and how does the annual inspection work?
A: [Company name] offers maintenance contracts for fire extinguishers, fire detection and automatic suppression systems. The annual inspection is performed by Apragaz-certified technicians and includes visual inspection, pressure measurement and refilling if needed. The inspection certificate is delivered digitally within 48 hours. Standard term: one year per BeNor NBN S21-050.
The difference: the good version contains concrete certifications, clear question phrasing, and a complete answer that's understandable without context. That version is citable. The bad version is not.
One warning with this example: the certifications, timeframes and standard references are placeholders. In your own FAQ, use only details that verifiably hold — an invented certification in a citable answer is precisely the hallucination you blame AI systems for, but home-made.