Define and test three customer personas

Turn a summary of your reviews, support tickets and surveys into three personas with AI, separate evidence from assumptions, and test the assumptions before you spend money on them.

Step by step

1. 1

   Start from your own data

   Ask an assistant to "create three personas for our customers" and you'll get convincing characters straight away. But it has never met your customers, so what you get is closer to a guess built from common stereotypes.

   A persona describes a group of customers with similar behaviour and needs, drawn as one person. Look for evidence of that behaviour and those needs in data you already have. Treat any opinion that doesn't come from customers themselves as an assumption until it's checked.

   Give the assistant counts and percentages rather than raw text. A review with the name removed can still be personal data if it can be matched back to the original — Korean law, for example, treats it as pseudonymised personal information. "Using AI at work safely" shows how to swap in fake values.

   Data that can back up a persona

   Source	What it tells you	Watch out for
   App or shop reviews	Customers' own words; what they like and dislike	Only some customers write reviews, so they don't speak for everyone
   Support tickets	Where people get stuck and what gets in the way	Mostly from people who ran into a problem
   Survey answers	Reasons and priorities you can count	Question wording and answer options change the answers
   Sales and usage data	What people actually do — when, how often, what	Shows what people did, not why
   Colleagues who talk to customers (support, sales)	Frequent questions and the mood on the front line	Their impressions are opinions — treat them as assumptions

2. 2

   Ask for personas with an evidence column

   Paste in your data summary and ask the assistant to note which numbers each persona rests on. With an evidence column, you can see straight away what to trust and what to check.

   This guide continues the 10-minute home workout app from "Draft A/B ad copy with AI". All the numbers are made up.

   We run a 10-minute home workout app for busy professionals. Here's a summary of the last three months, with personal data removed.
   - 120 app reviews (mentions per topic; a review can mention several): "before work / in the morning" 46, "too tired after work" 31, "worried about noise for the neighbours" 18
   - Of 80 support tickets: adjusting workout intensity 24, changing reminder times 16
   - Free-trial survey, 200 people, biggest reason for putting off exercise: no time 92, too tired 54, hard to stay motivated alone 38, other 16
   - Ages of 300 new subscribers: 20s 72, 30s 141, 40 and over 87
   Create three customer personas from this data as a table with the columns name, situation, goal, obstacle and evidence. In the evidence column, give the numbers from the data above that you used.

   I've split the data into three groups whose behaviour clearly differs. Percentages in brackets are out of the total for each data set.

   The persona table the assistant returned (example)

   Name	Situation	Goal	Obstacle	Evidence
   Early-bird Jake	In their 30s, office job, works out before work	Exercise every day, even briefly	No time; likes to sleep in	Reviews "morning" 46 (38%); survey "no time" 92 (46%); subscribers in their 30s 141 (47%)
   After-work Sarah	Worn out after work; short workouts in the evening	Keep going without overdoing it	Fatigue; workouts too intense for them	Reviews "tired" 31 (26%); survey "too tired" 54 (27%); tickets "intensity" 24 (30%)
   Quiet Mark	Works out alone at night in an apartment	Exercise without bothering the neighbours	Noise worries; motivation fades alone	Reviews "noise" 18 (15%); survey "alone" 38 (19%)

   Check it

   • Match every number in the evidence column against your summary. A number that isn't in your summary was made up. Work out the percentages yourself too (46 ÷ 120 ≈ 38%).
   • Note which persona has the thinnest evidence. Mark rests on just two numbers, and there's no telling whether those two (reviews and survey) come from the same people.
3. 3

   Separate evidence from assumptions

   Assistants fill gaps with details that aren't in your data, and those details read naturally. Asking for a label on every detail shows at a glance what needs checking.

   The easiest assumption to miss is the one that puts all the numbers into one person. The reviews, survey and sales data were collected separately, so these numbers alone can't tell you whether the morning exercisers are actually in their 30s.

   Break persona 1, "Early-bird Jake", into its details. Start each one with [Data] if the data I gave you supports it, or [Assumption] if not. For [Data], give the supporting number; for [Assumption], add one line on how it could be checked.

   [Data] In their 30s — 141 of 300 new subscribers (47%) are in their 30s.

   [Data] Works out before work, in the morning — mentioned in 46 of 120 reviews (38%).

   [Data] No time — 92 of 200 survey respondents (46%) named it as the biggest reason.

   [Assumption] Has an office job — only guessed from "before work". Add a question about work to the survey.

   [Assumption] Wants to exercise every day, even briefly — no data asks about goals. Check in interviews.

   [Assumption] Likes to sleep in — not in the data. Check in interviews.

   [Assumption] All of the above describe the same people — the three numbers come from separate data sets. Count the ages of users who work out in the morning.

   Check it

   • If more than half the details are assumptions, the persona is still a hypothesis. Four of Jake's seven details are assumptions, so test them in the next step before the persona shapes any ads or plans.
   • If the assistant labels something [Data] but the number isn't in your summary, change it to [Assumption].
4. 4

   Plan how to test each assumption

   For each assumption, decide how you'll check it and what result would confirm it — before you look. Set the bar afterwards and it's easy to read the results the way you hoped.

   Use a survey item or a data query for things you can count, and interviews for reasons and circumstances. In interviews, avoid leading questions like "Mornings are hard, aren't they?" Ask open questions about things that actually happened.

   A plan to test Early-bird Jake's assumptions (example)

   Assumption	How to check	Keep it if
   Has an office job	Survey item: "What do you mainly do on weekdays during the day?" (office or company job, self-employed, studying, caring for family, other)	"Office or company job" is the most common answer
   Wants to exercise every day, even briefly	Interview: ask what they hoped for on the day they downloaded the app	Several people bring up "every day" or "keeping it up" without being asked
   Likes to sleep in	Interview: on a day they missed a workout, ask what happened that morning	Several people bring up sleep or getting up before you ask
   All of the above describe the same people	Data query: ages of new subscribers who mostly work out between 6 and 9 a.m.	The share in their 30s is clearly above 47%, the figure for all new subscribers

   Write five interview questions to test these two assumptions.
   Assumption 1: They want to exercise every day, even briefly.
   Assumption 2: Liking to sleep in gets in the way of exercise.
   Ask open questions about things that actually happened. Leave out anything that can be answered yes or no, or that suggests an answer. Don't use the app's name or the word "sleep".

   Rather than asking about either assumption directly, these questions get people talking about things that happened recently.

   1. Think of the last day you exercised. When, where and for how long?

   2. On the day you downloaded the workout app you use now, what were you hoping for?

   3. If there was a day you meant to exercise but didn't, tell me what happened that day.

   4. Think of your most recent weekday morning. Walk me through what you did from the moment you got up until you left home.

   5. Tell me about one thing that has helped you keep exercising and one thing that has got in the way.

   Check it

   • Read each question aloud. If it can be answered with a plain yes or no, rewrite it.
   • Try the interview on a colleague first and fix any confusing questions. Tell participants what you'll record and how you'll use it, and get their consent.
5. 5

   Brief your ad copy with the checked persona

   Move confirmed assumptions to [Data], and fix or drop the ones that don't hold. Say the query shows that of 90 new subscribers who mostly work out between 6 and 9 a.m., 58 (64%) are in their 30s — well above the 47% overall. The fourth assumption then becomes [Data].

   Put the finished persona in the audience line of the brief from "Draft A/B ad copy with AI". Keep anything still being checked out of the centre of the copy.

   Product: a 10-minute home workout app for busy professionals
   Audience: people in their 30s who work out before work (58 of 90 new subscribers who mostly exercise between 6 and 9 a.m. are in their 30s)
   Evidence: 46 of 120 reviews mention working out before work or in the morning; 92 of 200 survey respondents put off exercise because they have no time
   Unconfirmed assumption (don't build a headline around it): they like to sleep in
   Channel: Instagram feed ad
   Goal: sign-ups for a 7-day free trial
   Write two headlines, each with one line on the evidence it relies on.

   1) "Ten minutes before you get ready, and today's workout is done" — relies on the 46 of 120 reviews that mention working out before work or in the morning.

   2) "No time to exercise? Ten minutes is enough" — relies on the 92 of 200 survey respondents who put off exercise for lack of time.

   Check it

   • Make sure neither headline leans on the unconfirmed assumption (sleeping in).
   • An A/B test tells you which line drew more response, not that the persona is right. Recheck the persona against fresh data every quarter.