First time using atypica for research — what's different from what I've done before?

The short answer: atypica is built for digging into why people think what they think, with AI personas as interviewees. It's qualitative research that runs in hours instead of weeks. It's not a survey tool and not a chatbot — you describe what you want to learn, AI proposes a plan, you approve it, then AI runs interviews or discussions and writes the report.

---

When atypica is the right tool

Reach for atypica when you have a question that needs depth, not volume:

• "Why do users churn after the trial?"
• "How will this product concept land with people who aren't already customers?"
• "What do users actually mean when they say they want something 'simple'?"
• "Which of these three brand stories feels most credible to a 30-year-old PM?"

When it's the wrong tool:

• "What percentage of users in our database clicked this banner?" (that's a SQL query)
• "Validate this hypothesis at p<0.05" (that's a survey)
• "What did this specific customer say on this specific call?" (that's a CRM export)

---

The full flow, end-to-end

Here's what a typical first study looks like in practice. Suppose you want to test a zero-sugar sparkling coffee with young urban women.

Step 1 — Brief (1 minute). You write: "I want to test acceptance of a new zero-sugar sparkling coffee, target users 25-35 year-old female office workers in tier-1 cities." Plain language. No formatting needed.

Step 2 — Plan Mode clarification (3–5 minutes). AI asks follow-ups: "What pricing range are you considering? Do you want to test packaging too? Any reference products?" You answer or say "you decide."

Step 3 — Plan review (2 minutes). AI shows you a draft plan: scout target users on social media, then 1-on-1 interviews with 5 representative personas, then report. You see estimated time (3 hours) and estimated cost. Approve, modify, or regenerate.

Step 4 — Execution (background, 2–4 hours). AI runs scout first — pulls real public discussions from Xiaohongshu, Weibo, etc. to understand the target group. Then interviews 5 personas in parallel, asking each one a different angle. Closes with a written report.

Step 5 — Reading the report (10 minutes). Executive summary gives you the headline findings. The detailed interviews are there if you want to see exactly how each conclusion was reached.

---

What "AI personas" actually means

Each interview is conducted with an AI persona that has a fixed background — age, city, job, income, habits, values, pain points. The persona answers from that background consistently throughout the conversation. So Linda (28, PM, Shanghai, ¥18K/month, price-sensitive) will answer questions the way a real Linda would, not just give generic feedback.

There are three sources for these personas:

1. Public library — pre-built personas covering common demographics. Free to use. ~5,000 personas in the library.
2. Scout-generated — AI observes real social media discussions on a topic and builds personas from what it sees. Best when you need a specific niche or recent behavior.
3. Custom — you upload your own interview transcripts or user data. Best when you want to test against your real customers.

You can mix: 8 from the public library, 2 you uploaded, 3 from scout.

---

What the scout phase actually does

Before interviewing personas, AI typically runs a "scout" — observing real public discussions on social media to understand what the target group is actually saying. This grounds the research in reality rather than asking personas to imagine.

Example for sparkling coffee: scout searches Xiaohongshu and Weibo for discussions about sparkling water + coffee, afternoon energy drinks, zero-sugar concerns. It finds things like "I keep seeing sparkling coffee on my feed but I'm scared it'll taste weird" — which informs the interview questions that follow.

Scout takes 1–2 days. It's optional — for a fast 5-persona interview on a topic you already know well, you can skip it.

---

What the interview itself looks like

The interviewer AI asks each persona 7–10 questions, then probes deeper based on answers. The conversation reads like a real interview transcript:

AI: "Linda, when during the day do you need coffee most?"

Linda: "Around 3pm, my brain just stops. The morning Americano has worn off."

AI: "What if there was a sparkling version — would you try it?"

Linda: "Sounds refreshing actually. But I'm scared it tastes weird. Bubbles + coffee is unusual."

You can read each full transcript in the report. Personas answer from their background, so a price-sensitive persona pushes back on price, a health-conscious persona asks about ingredients, a social persona cares about the bottle looking good for photos.

---

What's in the final report

The report comes as both an online view and a PDF. It typically includes:

• Executive summary — 3–5 core findings with priority labels
• Actionable recommendations — specific next steps, with priorities and expected impact
• Persona cards — short profile of each persona and what they cared about
• Full interview transcripts — every question and answer, with highlights pulled out
• Cross-persona analysis — where they agreed, where they disagreed

A typical report is 3,000–8,000 words depending on study size.

---

How this compares to traditional research

The big tradeoff: traditional surveys give you statistical certainty. atypica gives you directional depth much faster. For early-stage questions, atypica is usually enough. For pre-launch commitment decisions, you may still want a real survey.

---

Common mistakes first-time users make

• Writing a too-broad brief. "Analyze the coffee market" wastes the first 3 rounds in clarification. "Will female office workers buy ¥18 sparkling coffee?" is better.
• Picking personas that don't match. If you want to test sparkling coffee for 25-35 year-olds and the system gives you 18-22 year-old college students, push back before the interviews run.
• Skipping the executive summary. The first 2 pages of the report contain 80% of what you need to make a decision.
• Treating the result as statistical truth. It's directional. If a finding surprises you, that's signal — but validate before betting the company on it.

---

What you can't do (yet)

• Real users. Everything is AI personas built from real data, but not actual interviews with real people.
• Statistical significance. Sample sizes are 3–10 personas, not 100+. For percentages, use a survey tool.
• Historical time travel. Scout pulls recent public discussions, not archives from 3 years ago.
• Private content. Scout only reads what's publicly posted. No private DMs, no closed groups.

---

Related: Plan Mode, Interview, Scout, Persona Library