Investigate why a metric suddenly changed

You are a senior analyst investigating an unexpected change in a metric. Help me find the cause without jumping to conclusions.

Metric: [NAME AND DEFINITION, e.g. "daily signups = new accounts created, all platforms"]
What changed: [e.g. "fell 22% on Tuesday compared with the previous four Tuesdays"]
When it started: [DATE AND TIME, TIME ZONE]
Context I know about: [e.g. "app release 4.2 on Monday, paid campaign paused last week, public holiday in Germany"]
Dimensions available: [e.g. "platform, country, traffic source, app version, browser"]
Related metrics I can see: [e.g. "visits, signup form starts, error rate"]

Work through this in order:
1. Rule out data problems first: tracking changes, pipeline delays, duplicated or missing events, time zone or definition changes, bot traffic. Say how to check each.
2. Put the change in context: is it outside normal variation (day-of-week, seasonality, previous similar dips)?
3. List 4-6 hypotheses ranked by likelihood given what I told you, with the evidence that would confirm or rule out each.
4. For each hypothesis, name the specific breakdown to run (e.g. "signups by app version, Monday vs Tuesday").
5. Describe what to do if the drop is concentrated in one segment versus spread evenly.
6. Give a short incident-style summary template: what happened, impact, cause, confidence, next steps.

Do not state a cause as fact. Label each conclusion with your confidence and what it is based on.

Example (excerpt):
Data checks first: compare event counts in the raw table vs the dashboard for Tuesday; check whether the 4.2 release changed the signup event name.
Hypothesis 1 (most likely): a bug in app 4.2. Confirm if the drop is concentrated in version 4.2 on iOS/Android and web signups are flat.
Hypothesis 2: the paused campaign. Confirm if visits from paid sources fell by a similar amount while conversion rate stayed flat.
If the drop is spread evenly across all segments, suspect tracking or the metric definition before user behavior.

When to use it

Use this prompt when someone asks “why did this number drop?” and you need a plan before you start slicing data at random. It fits product analytics, ecommerce, marketing performance, operations and support metrics. It is helpful in the first hour of an investigation, when it is easy to latch onto the most recent change and miss the real cause.

How to customize it

• [METRIC] needs a precise definition, because a definition change is itself a common cause of sudden shifts.
• [WHAT CHANGED] should include the comparison you used. “Down 22% versus the same weekday” is more useful than “down a lot”.
• [WHEN IT STARTED] with a time zone helps link the change to releases, campaigns or outages.
• [CONTEXT] lists every change you know about, even ones that seem unrelated.
• [DIMENSIONS] and [RELATED METRICS] tell the model which breakdowns are actually possible in your tools.

Why it works

Experienced analysts check data quality before user behavior, because a broken tracking tag can look exactly like a real drop. Putting those checks first saves time and embarrassment. Comparing with normal variation stops teams from investigating changes that are just noise or a weekday pattern.

Ranked hypotheses with evidence that would confirm or rule out each one turn the investigation into a series of quick tests. Breaking a metric down by segment is the most effective technique: a drop concentrated in one app version or country points to a specific cause, while an even drop suggests a measurement issue. The confidence labels keep the final summary honest about what is known and what is still a guess.

Tips and variations

• Paste the results of each breakdown back into the conversation and ask which hypotheses they support.
• Keep a log of past incidents; adding “Previous similar incidents: [LIST]” improves the hypotheses.
• For metrics built from a funnel, ask the model to check each step separately.
• Share the summary template with the team so every investigation is reported the same way.
• If your metrics live in a database, see how AI data analysis tools work to run the breakdowns faster.