Article contents
A prompt panel is a sample, not a keyword list
A prompt panel is a sample of the questions your buyers actually put to generative engines, built so that its proportions reflect those of the real market. It is the basis for measuring citation share: every week, the same prompts are submitted to the same engines, and you record who is cited. If the sample is biased, so is the measurement, however rigorous the protocol.
The difference from a keyword list is not cosmetic. A keyword runs to two to four words and describes a topic; a prompt runs to one to three sentences and describes a situation: who is speaking, what they are looking for, under which constraints. The generative engine breaks that situation down into sub-queries and selects sources for each one. A keyword panel therefore measures a usage that buyers do not have, and it almost always overstates the visibility of generic pages.
Representativeness is judged on three dimensions: intent (what the buyer wants to obtain), language and market (where they buy), and stage of the journey (discovery, framing, choice, verification). A panel that covers all three well gives a readable citation share; a panel that neglects one produces a flattering or alarming figure with no connection to sales. The pillar article of this blog places this measurement within the overall method.
The most common bias is not technical: it is the panel that contains the questions the company would like to be asked, rather than the ones buyers ask. The former show off the offer; the latter contain constraints, hesitations and competitors' names.
Where to find real prompts
The best sources are those where the buyer phrases the question themselves, in their own words and with their own constraints. We rank them in decreasing order of fidelity.
- Sales teams and pre-sales. The questions asked in a first meeting, the objections and the requests for comparison are the prompts closest to the decision. A well-kept CRM record supplies dozens of them.
- Support, customer service and the site's chat. Tickets contain the situational questions (“I have this problem, how do I…”) that engines handle more and more often.
- Contact forms and internal search. The “message” field of a form is a goldmine of complete phrasings; internal search gives the raw vocabulary.
- Search Console, filtered on long question-shaped queries (how, why, which, how much). Short queries are not used as they are; they indicate topics, not prompts.
- Forums, communities and reviews, where peers phrase the question before buying, often with a budget and a context.
- The engines themselves. The follow-up questions that Perplexity or ChatGPT suggest after an answer, and Google's “People also ask” block, show the rephrasings the engine considers natural. They are used as a complement, never as the main source.
- Customer interviews, with one simple question: “What did you type or ask when you started looking?”
- 01CollectEvery source, unsorted, in a single file: raw phrasing, origin, date.
- 02CleanDeduplicate, anonymise, discard internal or off-market questions.
- 03Classify and balanceAssign intent, language, market, stage; apply the grid's quotas. This is where the panel becomes representative, or not.
- 04Rephrase in the buyer's voiceA situation, constraints, no SEO jargon, no brand name outside the brand intent.
- 05Double upA second phrasing for the decisive prompts, to absorb variability.
- 06Freeze and versionVersion number, date, an identifier per prompt; the panel only changes at the quarterly review.
Splitting by intent: the panel grid
The grid sets, for each intent, the share of the panel it occupies and what it measures. The proportions below are those of a typical B2B panel; they shift with the sector, but the order of magnitude for choice questions never drops below a quarter, because that is where citation influences the purchase.
| Intent | Target share | Example (publisher of invoicing software for SMEs) | What the cell measures |
|---|---|---|---|
| Informational | 30% | “How does mandatory e-invoicing work for an SME?” | Your role as an educational source |
| Comparative | 20% | “What is the difference between invoicing software and an ERP for a ten-person company?” | Your presence when the buyer frames their choice |
| Product or provider choice | 30% | “Which invoicing software should an SME choose when moving to e-invoicing?” | Your place on the shortlist |
| Brand | 10% | “Is [brand] any good for SME invoicing, and what are its weak points?” | What the engine says about you |
| Situational | 10% | “My accountant requires a structured invoice format, how do I do that without changing tools?” | The roundabout questions that lead to you |
Table: scroll horizontally.
Illustrative data. What to read here: the most frequent case in our audits is a panel that is two-thirds informational, which gives a high citation share and says nothing about the brand's place at the moment of choice.
Two practical remarks. The “brand” intent is measured with your competitors: the prompt “is [competitor] any good” belongs in the panel, because the answer often names you. And the “situational” intent is the hardest to collect, because it lives in support and in forums; it is also the one where competitor sites are least present.
Splitting by language and market
A language is a sub-panel, not a translation. Spanish, French and English-speaking buyers do not ask the same questions, do not cite the same competitors, and are not subject to the same regulation or the same seasonality. A prompt translated literally from French into Spanish produces a question no Spanish buyer would phrase, and the engine treats it as such.
Each sub-panel is therefore built from sources native to the market: local sales teams, local support, local communities. It follows the same intent grid, with its own quotas if the market requires it (a market in the discovery phase will weight informational questions more heavily). It receives a language and country identifier, because the same language behaves differently depending on the engine's interface country.
The reading is then done language by language. Engines answer in the language of the prompt but readily cite sources written in another language, especially English; a brand can be cited in Spanish thanks to its English pages and absent in French despite its French pages. Multilingual content and generative engines details this mechanism and its consequences.
What minimum size?
Size is derived from what you want to read, not from a round number. A reading cell (one intent in one language, for instance) becomes readable from about ten prompts; below that, a single prompt flipping moves the cell by ten points and nothing can be told apart from noise. A panel of thirty prompts per language therefore allows an overall reading and a reading across two or three grouped intents. A panel of fifty to a hundred per language allows the full grid to be read, engine by engine.
Beyond a hundred and fifty prompts per language, the weekly cost of measurement, with its three to five runs per prompt and per engine, grows faster than the information obtained. In our engagements, sixty prompts per language is the most frequent point of balance for an SME or a mid-sized company; a group present in several markets builds several panels of that size rather than a single giant one.
Write prompts the way your buyers do, not the way an SEO does
Phrasing changes which sources are retained. A short, nominal prompt brings back generic pages and comparison sites; a situated prompt, with constraints, brings back pages that answer those constraints. The panel must look like the second, because that is how buyers speak to an assistant.
Four writing rules follow from our records.
- A situation, not a topic. Who is speaking, what company size, what context, what time or budget constraint.
- No brand name outside the brand intent. A prompt that contains your name measures your awareness, not your visibility.
- No SEO vocabulary. Buyers say neither “best invoicing software 2026” nor “SaaS invoicing solution”; they describe their problem.
- Two phrasings for the decisive prompts. Choice and comparison questions receive a variant that changes the context without changing the intent. Both count as distinct prompts, in the same cell, and the gap between them tells you how sensitive the engine is to phrasing.
The vocabulary of the prompts then feeds the writing of the pages: aligning headings with real prompts starts with having a panel that contains them.
Maintaining the panel without breaking the series
A panel frozen forever becomes wrong; a panel modified as you go makes the series unreadable. The rule that reconciles the two is simple: a stable core, a controlled rotation, a dated version.
- Month 1Building version 1
- Multi-source collection, cleaning, classification
- Thirty to sixty prompts per language according to the grid
- Baseline citation share
- Months 2 and 3Stabilisation
- Spotting unstable prompts or prompts with no cited source
- Adding variants on the decisive prompts
- Version 1.1 frozen, series preserved
- Every quarterReview
- Rotation of at most 20% of prompts
- New questions from the field
- Four-week overlap between old and new prompts
Three practical details. The core, roughly 80% of the panel, does not move for twelve months; it is what carries the trend. Rotation replaces prompts that have become obsolete (regulation, withdrawn product), those that never produce a cited source because the engine answers from memory, and duplicates discovered in use. And each replacement is measured alongside the old prompt for four weeks, so that the series shows no unexplained break.
Finally, the panel is a shared document: sales teams add questions to it every quarter, marketing reads in it the intents that the pages lack, and management finds in it the terms its customers use. A panel that only the agency knows ages fast.
The most frequent panel mistakes
The panels we take over often show one of these seven weaknesses, and each one shows up in citation share before its cause is found.
- A panel exported from a keyword tool, without rephrasing: it measures a usage buyers do not have.
- The brand name present in the majority of prompts: the panel measures awareness and flatters the result.
- A panel that is two-thirds informational, with no choice or situational questions.
- Sub-panels translated word for word from one language to another.
- A panel modified without version or date: the series loses its comparability.
- A single phrasing per decisive question, which makes the measurement hostage to one word.
- A panel written by the person who writes the pages, who then measures their own writing.
What to remember
- A prompt panel is a sample of your buyers' real questions, not a list of rephrased keywords.
- The best sources are sales teams, support, forms, Search Console filtered on questions, and communities.
- The intent grid reserves at least a quarter of the panel for choice questions, where citation influences the purchase.
- A language is a sub-panel built from native sources; count on at least thirty prompts per language, sixty to read the full grid.
- A stable core of 80%, a quarterly rotation of at most 20% and a dated version preserve the measurement series.
Frequently asked questions
How many prompts does a panel need?
At least thirty per language for an overall reading, fifty to a hundred to read citation share by intent and by engine. The underlying rule is to have about ten prompts per cell you want to read. Beyond a hundred and fifty per language, the cost of the weekly runs grows faster than the information obtained; several medium-sized panels, one per market, are then preferable.
Can the panel be built from Search Console queries?
Partly. Long question-shaped queries (how, why, which, how much) are a good source, once rephrased as a complete situation. Short queries indicate topics, not prompts, and are not used as they are. Search Console also only covers Google and under-represents choice questions, which have to be sought from sales teams and support.
Should your brand name be included in the prompts?
Only in the “brand” cell, which represents about 10% of the panel and serves to check what the engine says about you, including the errors. Everywhere else, a prompt that contains your name measures your awareness, not your visibility: the engine cites you because it was asked to. Prompts such as “is [competitor] any good”, on the other hand, do belong in the panel, because their answers often name you.
How often should the panel be updated?
One review per quarter, with a rotation of at most 20% of prompts, and a core of 80% kept stable for twelve months. Each replaced prompt is measured alongside the old one for four weeks to avoid a break in the series. The panel receives a version number and a date at each review, and earlier measurements are kept with the version that produced them.
Do you need a separate panel per language?
Yes. Buyers in each market ask different questions, cite other competitors and are subject to other regulation; a prompt translated word for word is not a real prompt. Each sub-panel is built from sources native to the market, follows the same intent grid with its own quotas, and is read separately, because engines readily cite sources in a language other than that of the prompt.
An SEO practitioner since 2001, Kamel Malek runs SEO360 (Alicante, Valencia, Madrid, Paris). He has published three books on visibility in generative engines, including Generative Engine Optimization and Rétablir les faits.
