Connect the assistants

Bring your own keys. Each is sealed with AES-GCM before it is stored and is only ever unsealed inside the worker that makes the call. There is no read endpoint — once saved, a key can be replaced or deleted, never retrieved.

What are we measuring?

Give me a domain. I'll read the site and work out the brand, the product, who it competes with, and the topics buyers actually argue about. You get to correct all of it on the next screen.

Fix what it got wrong

This is the highest-leverage screen in the app. Everything downstream inherits these decisions, and a bad competitor list quietly corrupts twelve weeks of data.

Competitors

Topics

Keep names short — two or three words. They become axis labels in the 3D views and long ones clip.

Ground truth

Optional, and the highest-return five minutes in the whole setup. Paste what is actually true about you — prices, limits, features, guarantees. Every answer then gets checked against it, and contradictions come back as flagged findings instead of as sentiment.

The query set

These get sent verbatim, every run, forever. Once you start collecting, changing them breaks week-over-week comparison — so read them now. Roughly 60% deliberately do not name you: that is where you find out whether you get surfaced at all.

Collect

Each run is one snapshot. It executes on the server through a queue, so you can close this tab — the run keeps going and the progress is here when you come back.

Runs

Explore

Six ways into the same dataset. Coverage answers whether you showed up at all; the other five describe what happened when you did.