Most AI visibility work fails at the prompt list. The rest of the method (coding, engines, dates) cannot rescue a panel that only asks the questions a company already expects to win.
This brief is a design note for that list. It is written for researchers and research-minded operators who need a panel they can defend in a review meeting.
Write jobs, then write prompts
Start with the work a person is trying to get done. Then write the words they might actually type or speak.
Common jobs in commercial categories include:
- Orient: “What is this category and who is in it?”
- Compare: “How do two or three named options differ?”
- Select: “What should I use for a stated constraint?”
- Verify: “Is this claim or specification accurate?”
- Operate: “How do I do a task, or fix a failure?”
Each job produces different answer shapes and different citation habits. A panel that is 90 percent “best X for Y” questions is a selection study. That can be useful. It is not a complete visibility study.
For each job, write several prompts that vary the constraint (budget, region, company size, technical level, time horizon) without changing the job. Variation is how you learn whether a brand appears only in one narrow story.
Keep branded and unbranded prompts in separate ledgers
Unbranded prompts tell you whether the brand enters the category conversation. Branded prompts tell you whether the engine can identify the entity and describe it without mixing it up with a neighbor.
Both are useful. Mixing them in one average is not.
A simple split:
- Unbranded: category, comparison without your name, task how-tos, “alternatives to [competitor].”
- Branded: “what is [brand],” “is [brand] good for [job],” “[brand] vs [peer].”
If the unbranded set is empty, you do not have a visibility audit. You have a named-entity check.
Be careful with “alternatives to [competitor]” prompts. They are unbranded with respect to you, but they are not neutral. They inherit the competitor’s framing. Keep them, and tag them so they do not silently dominate the panel.
Cover the question types engines actually see
People do not only ask listicles of a model. They ask messy, specific, local, and follow-up questions. A panel that stays in polished marketing English will miss the way retrieval really gets stressed.
Include, in plain language:
- Short definitional questions.
- Comparison questions with two or three named peers (not ten).
- Constraint questions (“for a team that cannot install software,” “for a public-sector buyer”).
- Negative or troubleshooting questions.
- A few questions that should produce a refusal or a hedge, so you can see how the engine behaves when it is unsure.
You do not need hundreds of prompts to start. You need a set that you can run again. A first panel of a few dozen, mapped to jobs, is enough to learn whether the instrument works. Add prompts when a job is missing, not when someone wants a more flattering line.
Make the panel boring enough to reuse
Clever prompts feel like research. Reusable prompts are research.
Write prompts that a colleague could run next quarter without asking you what you meant. Avoid in-jokes, campaign slogans, and internal product names that customers do not use. If you must include an internal name, mark it as a branded identity check.
Give every prompt a stable ID, a job tag, a branded/unbranded flag, and a version. When you edit wording, increment the version and keep the old text. Quietly rewriting prompts is how time series die.
A collection record should capture:
- Prompt ID and version
- Engine and interface
- Date and time (with timezone)
- Visible settings (tools, browsing, location if shown)
- Raw answer, including sources as displayed
- Coder and codebook version
If you cannot produce that record, you cannot explain a change.
Watch for prompts that smuggle the answer
Some prompts are not questions. They are brand theater.
Examples of smuggling:
- Including the brand in an unbranded item.
- Using distinctive taglines as if they were category language.
- Asking “why is [brand] the leader” and then scoring the praise.
- Stuffing features into the question so the model has to mention them.
Those prompts can be used in a messaging test. They should not sit in the same table as a category audit.
A useful test: hand the unbranded list to someone who does not work at the company. If they can tell which brand paid for the study, rewrite the list.
How we use panels in Northline work
When we design a panel, we treat it as a research object that can be reviewed on its own, before any engine is queried. The review asks:
- Which buyer jobs are present, and which are missing?
- Is the branded set clearly separated?
- Are question types varied enough that one style cannot dominate the result?
- Can this panel be run again in three months without reinterpretation?
Only after that review do we collect answers. The panel is the claim about the world. The answers are what the engines did with that claim.
If you want help reviewing a panel, or want a panel designed for a category study, write to [email protected].