Realistic
Written the way customers write them, not the way marketing does.
Prompt research
Before measuring, choose what to ask.
Prompt research is choosing the questions to ask AI assistants in order to understand how they answer about your category and your brand. Find out where to find customers’ real questions and how to build a stable panel.
In brief
Prompt research: which questions to track.
Before measuring visibility in AI answers you have to decide what to ask. Prompt research builds a set of realistic, neutral and fixed questions: the panel on which brands and periods are compared.
Neutral
They do not suggest the brand in the question, unless needed.
Fixed
The same over time, so that comparisons make sense.
Where to find them
Real questions come from customers.
There is no public list of the questions people ask assistants. The best sources are those where customers already write their questions, in their own words.
Where a company finds customers’ real questions, what each source tells it and how to turn it into a panel question.
| Source | What it tells you | How to use it |
|---|---|---|
| Search Console | Long searches, often phrased as questions. | Keep those that describe a need, not a name. |
| Customer support | Doubts before and after purchase. | Rewrite frequent cases as general questions. |
| On-site search | What visitors who have already arrived are looking for. | Discover the words the site does not use. |
| Forums and reviews | How customers compare and recommend. | Take the comparison and choice questions. |
| Sales staff and partners | Objections heard in person. | Turn them into questions about features and price. |
Five types of question
Each type of question measures something different.
A category question tells you whether the assistant suggests you without anyone naming you; a brand question tells you how it describes you. A useful panel contains them all, in proportion.
Five types of question placed by closeness to the purchase decision and by whether the brand name appears in the question itself, with what each one measures.
Brand name in the question
Closeness to purchase
Problem
“How do I block out noise on a plane?”: measures whether you are suggested as a solution.
Category
“Best headphones for travelling”: measures whether you appear among the choices.
Comparison
“Northstar or a competitor?”: measures how it compares you.
Brand
“Is Northstar reliable?”: measures how it describes you and on which sources.
Support
“How do I update Northstar headphones?”: measures whether the facts are correct.
Writing well
A neutral question gives a readable answer.
The way a question is written changes the answer. Questions that suggest the brand or mix several requests produce measures that say nothing.
Four common mistakes in writing questions for AI assistants, each with a version that produces a comparable measure.
| Written badly | Written well | |
|---|---|---|
| Leading | “Why does Northstar make the best headphones?” | “Which headphones do you recommend for frequent travel?” |
| Too many requests | “Cheap, light headphones with good battery life, and where to buy them?” | One question per need, with a single main criterion. |
| Language | The same questions translated word for word in every market. | Questions rewritten in the words of each language and country. |
| Stability | Questions changed every month to chase what is new. | A fixed panel, with new questions added separately. |
From raw material to panel
Five steps to build the panel.
The panel starts from many raw questions and ends with a few dozen chosen ones, balanced across types and kept fixed to compare periods.
The steps by which questions gathered from customer sources become a balanced panel, written neutrally and reviewed at regular intervals.
Collect
Raw questions from search, support, forums and sales.
Group by need
Questions that ask the same thing become one.
Rewrite neutrally
Without the brand, where the question does not require it.
Balance the types
Problem, category, comparison, brand and support.
Fix and review
The same panel for months; new items go in separately.
Frequently asked questions
Frequently asked questions about prompt research.
How many questions you need, whether volumes of assistant questions exist, how prompts differ from keywords and how to handle several languages.
How many questions does a panel need?
A few dozen are enough to start, as long as they cover the question types and markets that matter. More questions give more stability, not more truth.
Are the volumes of questions asked to assistants known?
Assistant providers do not publish them. Search engine queries and customers’ questions remain the soundest indication of what people ask.
What is the difference between a prompt and a keyword?
A keyword is often short; a prompt is a complete question, with a need and a context. Many prompts come from long keywords rewritten as questions.
Should questions be translated for each market?
They should be rewritten, not translated: each language has its own words for the same need, and assistants answer differently from one country to another.
Related paths
After “Prompt research”, the next step.
The guides explain the concepts with examples on fictional brands. From here you continue with the other guides, the methodology and the features that apply them to your brand.
Guides to reputation, trust and AI
Guides on GEO, AEO, AI visibility, online reputation, reviews and monitoring: what to check, a worked example and the steps to act.
ReadFrom question to action
How TrustInsider works in six steps: brand, questions, quote, evidence, page-by-page intervention and a check on the same questions.
ReadA transparent method, not a black box
How every TrustInsider measure is built: sampling of AI answers, sources and rights, index calculation and declared limits.
ReadFind the right answer
Help with access, credits, collections, integrations and data: guides, FAQs, the site assistant and support requests from your workspace.
Read