With 300K AI Personas, aren't they all too similar?

Question Type

Product Q&A (TYPE-A)

User's Real Concerns

With so many personas, are they batch-generated using templates?
Are they just changing surface parameters like age and gender?
Will interview responses all be pretty much the same?

Underlying Skepticism

Doubt about persona diversity and differentiation

Core Answer

The answer is: No.

Behind the 300K personas are combinations of 7 dimensions, with theoretically astronomical combinations. More importantly, each persona is based on real user behavioral data, not simple parameter permutations.

Detailed Explanation

Why personas are NOT "all too similar"?

Reason 1: 7-Dimensional Three-dimensional Modeling

Not simple "age + gender + occupation", but deep characterization across 7 dimensions:

Dimension	Not	But Instead
Demographics	❌ 28-year-old female	✅ 28 years old, monthly income ¥15K, internet product manager, bachelor's degree
Geographic	❌ Shanghai	✅ Renting in Xuhui District, Shanghai, 1-hour commute
Psychographic	❌ Values quality	✅ Willing to pay for good products, likes trying new things, but worries about wasting money
Behavioral	❌ Online shopping	✅ Primarily online shopping, likes reading reviews, compares multiple brands, trusts Xiaohongshu (RED) recommendations
Pain Points	❌ No time	✅ Busy with work, no time to browse stores, worried about buying inferior products, wants quick decision-making
Technology Adoption	❌ Likes new tech	✅ Early adopter, willing to try new apps, interested in AI products
Social Relationships	❌ Single	✅ Single, friend circle mostly same-age white-collar workers, often shop and dine together

Same "28-year-old Shanghai female product manager," but completely different personas due to other 6 dimensions:

Persona A (Zhang Li):

Psychographic: Pursues cost-effectiveness, worries about wasting money
Behavioral: Compares multiple brands before online shopping, trusts Xiaohongshu (RED)
Pain Points: Busy with work, no time to browse stores
→ Reaction to "sparkling coffee": "Over ¥30 is too expensive, my coffee budget is usually ¥15-20"

Persona B (Li Yue):

Psychographic: Values quality and experience, willing to pay premium for good products
Behavioral: Buys expensive items directly, doesn't like comparing prices
Pain Points: Worried about buying inferior products affecting health
→ Reaction to "sparkling coffee": "¥35-40 is acceptable, but need to check ingredients, can't have too many additives"

Reason 2: Based on Real Users, Not Templates

Template generation problem:

Template: {age}-year-old {gender} {occupation}, monthly income {income}Generated: 28-year-old female product manager, monthly income ¥15K          29-year-old female product manager, monthly income ¥18K          27-year-old female product manager, monthly income ¥12K

→ This generation method leads to personas "all too similar"

atypica's construction method:

Real data sources:- Xiaohongshu (RED) User A: Posts discussing "cost-effective coffee", focused on price- Xiaohongshu (RED) User B: Shares "specialty coffee shops", focused on experience- Xiaohongshu (RED) User C: Shows off "trendy coffee", focused on photosPersonas built from real behaviors:- Persona A: Price-sensitive type- Persona B: Quality-first type- Persona C: Social sharing type

→ Each persona corresponds to real user behavioral patterns

Reason 3: Consistency Validation Ensures Real Differences

How do we verify personas are truly different?

Test method:

Select 10 "28-year-old Shanghai female product manager" personas
Ask the same question: "Would you buy ¥30 sparkling coffee?"
Observe if responses show real differences

Result examples:

Persona	Response	Core Focus
Zhang Li	"Too expensive, my coffee budget is usually ¥15-20"	Price-sensitive
Li Yue	"If ingredients are good, ¥35-40 is acceptable"	Quality-first
Wang Yue	"If packaging looks good, I'd buy it for photos"	Social attributes
Chen Si	"Would care about artificial sweeteners and additives"	Health ingredients
Zhao Xin	"Can try occasionally, but won't buy regularly"	Limited budget

Key Findings:

✅ Even with same basic information, focus points are completely different
✅ Response logic and values reflect real differences
✅ Not surface parameter differences, but deep decision-making logic differences

Real Case Comparison

Case: Same "25-30 year-old fitness enthusiasts"

Template generation problem:

Interview 5 "25-30 year-old fitness enthusiasts":

Persona 1: "I like fitness"
Persona 2: "Fitness is good for health"
Persona 3: "I work out 3 times a week"
Persona 4: "Fitness is important"
Persona 5: "I think fitness is good"

Problem: All feedback is similar, can't differentiate

Built from real data:

Interview 5 "25-30 year-old fitness enthusiasts":

Persona 1 (Li Ming, muscle building focus): "I mainly work out to build muscle. Go to the gym 5 times a week, each time training one body part. I strictly control protein intake, at least 150g daily. For fitness apps, I care most about exercise library and training plans."

Persona 2 (Zhang Yue, fat loss focus): "I work out to lose weight. Go to the gym 3 times a week, mainly cardio. I use apps to track calories, keep it under 1500 daily. Most concerned about syncing weight data and seeing progress curves."

Persona 3 (Wang Hao, social focus): "I mainly work out to make friends. Go to group classes 2-3 times a week, like spinning and Pilates. I like fitness apps with social features, can see friends' check-ins, encourage each other."

Persona 4 (Chen Si, rehabilitation focus): "I started working out because of back injury, mainly rehabilitation training. 2-3 times a week, movements must be very precise. I need apps that can guide proper form, preferably with video demonstrations, to avoid re-injury."

Persona 5 (Zhao Xin, habit formation): "I work out to form long-term habits. Don't care what I practice, key is consistency. I use apps to check in, seeing consecutive days gives me achievement. Most afraid of complex plans, I give up easily."

Comparison Results:

✅ Same "fitness enthusiasts," but completely different motivations
✅ Completely different needs for fitness apps
✅ Interview insights can guide product design

Data Support

Theoretical Space of 7-Dimensional Combinations

Simplified calculation (reality is more complex):

Dimension	Typical Value Count	Example
Demographics	100+	Age(5 brackets) × Gender(2) × Occupation(20+) × Income(5 brackets)
Geographic	50+	Tier 1-3 cities × Urban/Suburban
Psychographic	20+	Value and lifestyle combinations
Behavioral	50+	Consumption habits, decision patterns
Pain Points	30+	Needs and concerns in different areas
Technology Adoption	5+	From resistant to early adopter
Social Relationships	10+	Single/Married × Social circle type

Theoretical combinations ≈ 100 × 50 × 20 × 50 × 30 × 5 × 10 = 150 billion+

Practically meaningful combinations ≈ Tens of millions (excluding extreme unreasonable combinations)

Current library 300K = Covers 0.3-1% of common research scenarios

How to Verify Personas Are NOT "all too similar"?

Method 1: Consistency Test

Same persona, multiple responses to same question:

Public personas: Stable consistency
If "all too similar," consistency should approach 100% (completely mechanical)
Real persona characteristic: Both stability and humanized randomness

Method 2: Diversity Test

Different personas, same question:

If "all too similar," responses should be highly similar
Actual test: 10 "28-year-old female product managers" responses, average similarity only 30-40%
Diversity reflected in values, decision logic, focus points

Bottom Line

"300K personas are not 300K template copies. They are 300K real user behavioral patterns and decision-making logics."