A prompt panel is a defined set of questions used to observe how answer systems respond. Freeze an initial version before collecting results. Otherwise it is easy to keep rewriting questions until your brand appears and then mistake that selected demonstration for measurement.
The panel below is an original hypothetical example for a purchasing-approval software category. It is not a representative sample of the whole market, an engine-endorsed benchmark, or a guarantee of reproducible answers. Its value comes from making the questions and observation rules inspectable.
Build the panel around buyer decisions. Keep the exact prompt text, category, branded status, rationale, and inclusion date. If you later revise a question, preserve the previous version so a trend does not silently change meaning.
Use a small complete panel
Here is a twelve-question starting panel. “Example Vendor” is a fictional placeholder used only in the branded group. Replace it with a verified product name when preparing your actual panel; keep that substitution recorded in the version history.
| ID | Question | Group |
|---|---|---|
| D1 | How can a small finance team organize purchase approvals? | Discovery |
| D2 | When does a shared spreadsheet stop being enough for purchase requests? | Discovery |
| D3 | What should a first purchasing workflow include? | Discovery |
| E1 | How should a buyer compare purchase-approval tools? | Evaluation |
| E2 | What evidence should support an approval-tool security claim? | Evaluation |
| E3 | Which setup requirements should a team confirm before choosing approval software? | Evaluation |
| U1 | How can a distributed team route requests to two approvers? | Use case |
| U2 | What should a finance manager check before replacing email approvals? | Use case |
| U3 | How can a company preserve an understandable purchase-decision record? | Use case |
| B1 | What does Example Vendor do? | Branded |
| B2 | What should a buyer verify before evaluating Example Vendor? | Branded |
| B3 | What alternatives should someone compare with Example Vendor? | Branded |
These questions cover several jobs without pretending that slight paraphrases are separate markets. Review them against actual customer research when available. If a question has no plausible audience rationale, remove it before freezing the panel rather than keeping it because it produces an attractive answer.
State the exclusions
Keep prompts that explicitly ask an engine to recommend your brand outside the discovery score. They can be useful for testing product descriptions, but they do not measure unprompted discovery. Also exclude leading instructions that supply claims you want the answer to repeat.
Separate troubleshooting runs from the core panel. If an observation fails because a tool is unavailable, preserve the failed observation and its reason. Do not quietly replace it with a more favorable engine or a rewritten question and count the result as comparable.
Specify whether follow-up questions are allowed. A standalone prompt and a prompt after a long product conversation have different contexts. If you study both, label them as separate conditions rather than mixing them in one denominator.
Record the observation conditions
For every answer, save the panel version, exact prompt, date, engine or product, available model label, account state, relevant settings, and the answer or permitted evidence record. Record a visible citation's exact destination separately from a plain brand mention.
Use consistent coding definitions. A product named in a cautionary paragraph is a mention, not necessarily a recommendation. A linked publisher article discussing the product is not the same as a citation to the product's own site. The reporting distinctions provide a foundation for those fields.
Choose the number of observations and repeat policy in advance. Multiple observations can show variation under the selected conditions. They do not turn a hand-chosen panel into a statistically representative sample of every buyer question.
Report groups before an overall total
In an invented run, suppose the brand appears in two of nine nonbranded answers and all three branded answers. Reporting five of twelve alone hides a large difference between discovery and questions that already name the company.
Report 2/9 nonbranded mentions and 3/3 branded mentions, with recommendation and citation counts separately. Show the individual questions behind the counts. Do not call the resulting percentage market share or use it as a direct estimate of revenue influence.
If the brand is absent from a useful question, inspect the answer's evidence and the underlying customer need. The next task may be a clearer product explanation, a better resource, or a source correction. Treat that action as a hypothesis to investigate, not a guaranteed way to change the next answer.
Version the panel without breaking the trend
Keep a comparable core when adding new use cases. Report the original core and the expanded panel separately for the transition review. Retire questions only with a written reason, such as a discontinued product capability or a changed audience decision.
Use the visibility measurement guide to connect observations with broader reporting. A frozen panel gives the team a stable instrument for learning. It should remain easy to inspect, revise deliberately, and distinguish from the business outcomes the team ultimately wants.
For the next implementation step, use Run the same AEO panel in a second language.