Friction investigator

The friction investigator is an AI assistant attached to each friction finding. Ask it about the people who hit the problem, and it runs real queries on your data and shows you the evidence behind every number.

Updated

What it is

A friction finding tells you what happened. The investigator helps with the next questions. Who were these people? Where did they come from? What did they do instead? Every conversation is scoped to the people who hit that finding on that page.

Ask a question

  1. Open a finding under Friction.
  2. In Ask about this, type a question or pick a suggestion, such as Which plans were these people on? or Where did they come from?
  3. The panel Ask about this finding opens, Scoped to the 312 people who hit this on /checkout, and the answer streams in.
  4. Ask follow-ups in the same panel. Each builds on the conversation so far.

What it can look at

The investigator can only read. It has a fixed set of tools:

  • the finding’s own details
  • the events behind it (up to 50 rows per look)
  • your saved funnels
  • your saved segments, and Overview’s page, referrer and country breakdowns
  • one person’s profile and activity
  • a pre-filled experiment setup link, which it offers but never creates

It can’t change settings, create reports, or see anything outside your project. It treats the contents of your events as data, never as instructions.

Evidence and honesty

  • Every answer lists Evidence this answer used: chips that open the exact events, settings or evidence, with the time window each covers.
  • Every number in an answer comes from a query it ran. When the data can’t settle a question, it says so rather than guessing.
  • It is careful about why. It tells you where a problem is concentrated, not what caused it.
  • Every answer is labelled AI-generated — verify against the linked evidence. Open the chips before you act on an answer.

Limits

  • 100 questions a month for each project, reset at the start of each month (UTC). The panel shows how many are left.
  • 20 questions per conversation. Conversations aren’t saved yet: leaving the page starts a new one.
  • A very broad question may hit the per-question budget. The answer then suggests asking a narrower follow-up.

Good questions

Is this concentrated in one group?

  1. Ask: Is this concentrated in a saved segment?
  2. Then: Which browsers and devices were these people on?
If most affected people share one browser, you likely have a bug to reproduce, not a design problem.

Did it cost us conversions?

  1. Ask: What happened after people hit the Apply coupon button?
  2. Then: How did these people do in the Checkout funnel?

Show me one real example.

  1. Ask: Show me one person who hit this and what they did around it.
  2. Open the evidence chip to read their full profile in People.