About
Built to test one thing: does an insurance brief match what car owners write?
Why it exists
Good products start from understanding the customer. Many briefs start from a hunch, and teams build features nobody asked for.
On 2026-10-07 we read a brief from Tasco Insurance, a Vietnamese insurer that also runs a toll-payment app. It listed six beliefs about its customers. We wanted to know whether those are the problems car owners write about most. So we read what happens to car owners, from buying insurance to the day of a crash.
Atlas Insight is what came out. A team can test a belief against real customer talk in minutes. It can cut a wrong feature before it costs money, and find needs nobody thought to ask about.
How it was built
- We collected 14.5 million public posts from three car-owner forums, 2003 to 2026.
- We placed 13.5 million of them on a topic map (we call these analysed) and tagged about 10%, the insurance ones, in detail: topic, insurer and incident.
- We separated car owners from sellers, ads and copies.
- We checked the tags by reading samples by hand: about 8 in 10 are correct.
- We serve it two ways: a map to look around, and a connector so your AI assistant (ChatGPT, Claude or similar) can search all of it.
| Measure | Number | Since | Scope |
|---|---|---|---|
| Posts collected | 14,531,108 | 2003-09 to 2026-10 | public car-forum posts, all authors |
| Posts analysed by topic | 13,518,238 | 2026-10 | every post placed in a topic group |
| Insurance posts by car owners | 76,236 | 2003 to 2026 | sellers and ads removed |
| Posts tagged in depth | about 10% | 2026-10 | insurance focus |
| Topic tags correct | 8 in 10 | hand check, 2026-10 | posts read by hand against their tags |
| Questions answered with counts and posts | 50 | 2026-10-15 | two sessions of 25 questions each |
| Brief beliefs tested | 6 | 2026-10-15 | 4 did not hold, 2 barely mentioned, 0 hold |
| Outside teams using it | 0 | 2026-10-15 | none yet |
Known limits
Insurance posts by car owners peaked at 8,487 in 2012 and were 619 in 2025. After 2020, much car talk moved to social apps. Younger car owners trust forums less. We have not collected social apps yet.
What comes next
- Analyse the remaining 90% of posts.
- Test collecting posts from social apps.
- Let AI assistants search the posts for patterns nobody asked about.
- Check answers automatically by reading the posts behind them.
- Give AI assistants background material: laws, insurer rules, market figures.
Data, consent and how to connect
| Source | Public posts on three car forums, September 2003 to October 2026. Every post keeps a link to its source page. |
|---|---|
| Personal data | We collect no accounts, phone numbers or other personal data. These pages carry no tracking. |
| The map | Shows excerpts of posts, each linked to the original. |
| Hosting | We host it. A team needs no customer data to use it and sends none. |
| How to connect | Your own assistant (Claude, ChatGPT, Cursor) connects through a standard connector called MCP. It can only read, and answers about 30 questions a minute. |
| Not done yet | No login, one server, and no outside security review yet. |
Who builds it
vad builds and runs it alone; the first version took 11 days. To try it, connect your AI assistant (two minutes). Questions or data requests: