Investigate a sudden metric spike or drop

Systematically diagnose why a metric suddenly changed, ranking likely causes and giving the exact checks to confirm each one.

Change note: Initial version

Prompt

A metric changed unexpectedly and I need to find out why. Act as a methodical data investigator.

Metric: {{metric}}
What happened: {{change}}
When it started: {{timing}}
What I know about the business, product, or system around that time: {{context}}
Dimensions I can break the data by: {{dimensions}}

Produce an investigation plan:
1. Restate the anomaly precisely: size, shape (step, spike, slow drift), and timing, and what it should be compared against (same weekday last week, same period last year).
2. First, rule out measurement problems: tracking or pipeline changes, delayed or duplicated data, timezone or definition changes, bots. Give the checks in order of cheapness.
3. Then list 6-8 candidate real causes, ranked by prior likelihood given my context. For each: the evidence that would support it, the evidence that would contradict it, and the exact breakdown or query to run.
4. A decomposition approach: split the metric into its components (for example volume x rate, or by segment) and show how to find which component explains most of the change.
5. A decision tree: "If X is true, go to Y."
6. What to tell stakeholders right now while the investigation is ongoing, in 3 sentences, without claiming a cause I have not confirmed.
7. Follow-up: an alert or monitor to catch this sooner next time.

Say what extra information would most narrow down the cause.

Variables

• {{metric}}Metrictext

  Default: Daily signups on the website

• {{change}}What changedtext

  Default: Dropped about 35% overnight and has stayed lower for five days

• {{timing}}Whentext

  Default: Started Tuesday morning, 9 days ago was normal

• {{context}}Contextlong

  Default: We released a redesigned signup form on Monday evening. Marketing paused one ad campaign last week. No known outages.

  Releases, campaigns, outages, holidays, anything that changed.

• {{dimensions}}Available breakdownstext

  Default: Device, browser, country, traffic source, signup step funnel

Example output

**1. The anomaly**: A step change (not a spike or drift), about -35%, starting Tuesday morning and persisting five days. Compare with the same weekdays of the previous three weeks, since signups are likely weekly-seasonal.

**2. Measurement checks (cheapest first)**
- Is the signup event still firing on the new form? Compare front-end events with database user creation counts.
- Did the dashboard filter or metric definition change?
- Did bot filtering or duplicate suppression change in Monday's release?

**3. Candidate causes**
1. Redesigned form broken on a browser or device. Supports: drop concentrated in one segment. Contradicts: uniform drop. Check: signups by browser and device, before vs after.
2. Added friction in the new form. Check: funnel completion by step; a sharp drop at one field points to it.
3. Paused ad campaign. Check: signups by traffic source; if the lost volume matches that campaign's old volume, it explains the drop.
4. Tracking missing on the new form (measurement, not real). Check: database counts vs analytics.
5. Country-specific issue, such as a payment or validation rule. Check: signups by country.
6. Seasonality or a holiday. Check: same period last year.

**4. Decomposition**: signups = visits x visit-to-signup rate. If visits are flat, the rate fell; segment the rate by device and funnel step.

**5. Decision tree**: Visits down -> traffic source breakdown. Visits flat, rate down everywhere -> form change. Rate down in one browser -> bug. Database flat but analytics down -> tracking.

**6. Stakeholder note**: "Signups are down about a third since Tuesday. We are checking whether it is a tracking issue, a form problem, or lower traffic. We will update you by end of day."

**7. Follow-up**: alert when daily signups fall more than 20% below the same-weekday 4-week average.

**Most useful extra info**: signup counts from the database vs analytics, split by device and browser.

Notes

Ruling out measurement problems before hunting for real causes avoids the most common waste of time. Add a screenshot description of the trend or sample numbers to sharpen the ranking.