diff --git a/docs/root-cause-analysis.mdx b/docs/root-cause-analysis.mdx index 51fe5574..a17dd2b9 100644 --- a/docs/root-cause-analysis.mdx +++ b/docs/root-cause-analysis.mdx @@ -13,13 +13,17 @@ Mixpanel's **AI-Powered Root Cause Analysis (RCA)** feature does that work for y When you run an RCA, Mixpanel immediately launches a new Board and exposes the agent's reasoning as it works: -1. **Validates the anomaly.** Confirms there's a real change to explain before going further, using a z-score to check the change is well outside normal day-to-day variation rather than routine noise or seasonality. +1. **Validates the anomaly.** Confirms there's a real change to explain before going further. RCA compares your flagged point against a baseline window that ends right before it, sized automatically from your chart's time grain (hourly uses roughly the prior 24 hours, daily ~14 days, weekly ~9 weeks, monthly ~6 months); if your chart shows a longer range than that, RCA uses your range instead. It then measures how many standard deviations your point sits from that window's average (the z-score), with 2 or more being the bar for a real anomaly rather than routine noise. A seasonality check compares like-for-like periods (same weekday for daily data, same hour-of-day for hourly) so normal recurring patterns don't get flagged. 2. **Selects properties to analyze.** An LLM agent picks which properties to break down on, combining properties Mixpanel suggests for the event with any you've prioritized in [Custom RCA Settings](#custom-rca-settings). It spreads picks across plausible cause categories (acquisition, audience, platform, data quality, user identity), filters out high-cardinality identifiers and near-duplicates, and factors in your project's business context. 3. **Runs breakdowns.** Executes the chosen breakdowns to see which segments actually moved. 4. **Writes an interpretation.** Fills the Board with contribution factor analysis (each segment's share of the total change), ranked dimensions, an AI-written explanation of the likely cause, a confidence level, and suggested next steps. Cards stream into the Board as each step completes, with a live progress banner at the top while the agent is still working. + + **What the numbers mean.** The percentage change (sometimes called "lift") compares the data point you flagged against a baseline, which is the average of the periods just before it: `(flagged value - baseline) / baseline`, so +30% means it came in 30% above the recent norm and a negative number means below. Each segment's contribution percentage is its share of the total change: `(how much the segment moved) / (how much the whole metric moved)`, so if a metric dropped by 500 overall and one country accounts for 400 of that, it shows an 80% contribution. Confidence works differently: it's not a calculated number but a label ("High," "Medium," or "Low") the AI assigns to each finding based on how unusual the change is versus normal fluctuation (the z-score) and how cleanly a single segment explains it, so a big, clear change driven by one segment earns higher confidence than a small change spread across many segments or one that looks seasonal. + + ## Where to launch RCA