When your business serves different buyers, you need to know which of them see your brand recommended for their particular needs. Assistants answer conditionally. Supply a different description of who is asking and the same question returns a different set of names. Comparing across personas measures which buyers you are present for.
Set up the comparison
A campaign holds three personas besides a neutral one, each described in 10 to 50 words in the first person. The model answers as that person, so the constraints you name are what the answer responds to.
Spend the slots on buyers you would give different advice to. If two would get the same recommendation for the same reasons, they are one persona and you have a slot free.
A useful persona names what the buyer is responsible for, what they lack, what limits the decision and what would rule an option out. Job title alone rarely changes an answer.
The four patterns
| Pattern | Reading | Work |
|---|---|---|
| Present for all personas | Established across the market's written record | Watch for change; spend effort elsewhere |
| Present for one only | Your positioning landed with one buyer type | The material that achieved it exists and can be adapted |
| Absent for the buyer you sell to most | The most actionable finding available | Content and placement work from the cited sources |
| Present only without persona context | You appear when no constraints are applied and drop out when real ones are named | Your material says what you do, not who it is for |
The fourth row is why this exercise exists. It is common, it is invisible without persona tracking, and it has a specific fix: add information about fit to material that describes only capability.
Reading it fairly
Answers vary between runs regardless of persona, so compare across a series rather than reacting to one contrast. A difference visible in one run of each is not evidence of a difference between personas.
Keep the prompt wording identical across personas. If the wording differs, you are comparing answers to two questions.
The cost
Each persona variant multiplies collection. Ten prompts across three providers is 30 calls per ticked day; the same ten across three personas is 90.
Do not multiply personas beyond the ones that change the answer. Two well-chosen personas tracked properly beat four tracked thinly.
What a persona result is not
It is not evidence about real people. Nobody with that job has been surveyed. What is measured is how an assistant answers when given that description.
That is still useful, because it is the same mechanism your buyers encounter. It is not market research, and a persona finding should not be reported as what customers think.
Use the persona comparison when choosing which buyer group to investigate. Read the relevant answers before deciding what content or coverage that group needs.
See writing personas that change the answer, campaigns, or LLM Top Sites.