What data sources does AI Research use?
The short answer: two separate sources, depending on the step. Social Media Scanning pulls from real public posts on Xiaohongshu, Douyin, X/Twitter, Instagram, and TikTok. Interview and Discussion use atypica's pre-built persona library (~5,000 personas), which doesn't depend on social media data — those personas were built from a mix of real interview transcripts, user profiles, and prior Scout runs.
Sources by feature
| Feature | Data source | Privacy scope |
|---|---|---|
| Social Media Scanning | Public posts on 5 platforms | Public only — no DMs, no closed groups |
| Interview | Pre-built persona library | Built from anonymized research and Scout outputs |
| Discussion | Pre-built persona library | Same as Interview |
| Scout-generated personas | Public posts (above) | Same as Scanning |
You don't pick one source — different features use different sources automatically.
What Social Media Scanning actually pulls from
Domestic (China): Xiaohongshu, Douyin.
Overseas: X/Twitter, Instagram, TikTok.
Coverage is recent public posts only — typically the last few weeks to months. No historical archives. No private accounts. No login-gated content.
What Interview/Discussion personas are built from
The public persona library has ~5,000 personas covering common demographics. Each persona is built from a mix of:
- Real user interview transcripts (anonymized)
- Behavioral profiles from research contexts
- Outputs from prior Scout runs
Personas are stable: once a persona exists, it answers consistently across all interviews and discussions. Personas don't update their behavior based on real-time social media — they speak from their fixed background.
Why Interview doesn't check social media in real-time
If you ask a persona in an Interview "would you buy ¥18 sparkling coffee?", they answer from their persona background — not from what's trending on Xiaohongshu today. This is intentional:
- Personas are consistent, so research is reproducible
- Real-time social data wouldn't fit a single persona's perspective
- Mixing live data with persona responses would muddy what the persona is supposed to represent
For real-time trend data, use Scout. For depth on individual motivations, use Interview. The two complement each other.
Choosing platforms for Scout
Let AI pick by default — it matches platforms to your demographic.
If you want to specify manually:
| Demographic | Best platform(s) |
|---|---|
| Young women (18-35), consumer goods | Xiaohongshu |
| Gen Z (18-25), short video | Douyin |
| Lower-tier market | Douyin |
| Tech, business, opinion leaders | X/Twitter |
| Fashion, lifestyle, visual brands | |
| Young global users, product reviews | TikTok |
For cross-validation, scan 2–3 platforms and look for consistency. Contradictions signal you need more digging.
Regional filtering
Scout accepts regional filters:
- Tier-1 cities (Beijing, Shanghai, Guangzhou, Shenzhen)
- New tier-1 cities (Chengdu, Hangzhou, Wuhan, etc.)
- Provinces (Sichuan, Zhejiang, etc.)
- Overseas regions (USA, Europe, etc.)
Filter too narrow and there's not enough content. Filter too broad and the output gets noisy. Match your filter to your actual research question.
Data freshness
| Source | Freshness |
|---|---|
| Scout | Recent (last few weeks to months) |
| Persona library | Built periodically, not live |
If your research question is about current trends or recent events, run Scout first. If your question is about stable attitudes or decision-making patterns, Interview alone is enough.
What isn't included
- Historical archives — Scout only pulls recent posts. No "what were people saying in 2023" research.
- Private DMs and closed groups — only public posts are scanned.
- Real-time social listening — Interview/Discussion personas don't know what's trending today.
- Cross-platform identity resolution — Scout can't reliably link the same person's posts across Xiaohongshu and Weibo.
For any of these, supplement with a dedicated social listening tool (e.g., specialized analytics platforms).
Privacy and compliance
What gets scanned: publicly posted content only. No private messages, no closed-group discussions, no login-required pages.
What's extracted: themes, attitudes, language patterns, behavioral observations — not user IDs or identifying details.
What stays out: original post text isn't stored after the run; only summarized insights are kept.
Related: Scout, Interview, Discussion, Persona Library