Daily active users dropped 15% week over week. How do you diagnose it?

Do not start with hypotheses

The instinct is to start guessing: a bad release, a competitor, a holiday. Resist it. Interviewers score this question on process, and an unstructured list of causes — however plausible — reads as someone who would flail in a real incident. The structure below is what they are listening for.

Also, ask what the metric means before you diagnose it. "Active" is a definition someone chose: app open, or a specific action, or a session over some duration. Several apparent metric drops turn out to be a changed definition, and asking is a cheap way to look like you have done this before.

Step 1: is the drop real?

Instrumentation is a more common cause than behaviour, and it is fast to eliminate.

Check whether a release changed or removed the tracking event, whether the analytics SDK was upgraded, whether a data pipeline job failed or ran late, and whether a consent or privacy change is now suppressing collection for a subset of users. Check whether other metrics moved consistently — if DAU fell 15 percent but revenue, server request volume, and support ticket volume are all flat, you are looking at a measurement problem, not a user problem. That cross-check is the single most valuable thing you can say early.

Also confirm the comparison is fair. Week over week against a week containing a public holiday, or a week that included a one-off marketing spike, will show a "drop" that is an artefact of the baseline.

Step 2: is it broad or concentrated?

This is the question that determines everything downstream, and it has only two answers with very different implications. A drop spread evenly across all segments points at something global — a pricing change, an outage, a policy change, seasonality. A drop concentrated in one segment points at something specific and usually much easier to fix.

Cut along these dimensions, in roughly this order:

Platform and version. iOS versus Android versus web, and by app version. A drop confined to one platform after a release is nearly always that release. This is the highest-yield cut, so do it first.

Geography. A single country or region suggests a local outage, a regulatory or app-store change, a payment-provider failure, or a local competitor promotion.

New versus returning. If new-user activity is flat but returning users dropped, the problem is retention or an experience change. If returning is flat and new dropped, the problem is acquisition — a paid campaign ended, a channel's spend was cut, or organic ranking moved. These are different teams and different fixes, so the cut is decisive.

Cohort. Look at whether a specific signup cohort is churning faster, which points at something that changed in onboarding.

Acquisition channel. A drop confined to one channel is usually a marketing or attribution change rather than a product one.

User segment by engagement. Losing casual users and losing power users are very different problems even at the same magnitude, and the second is far more serious for revenue and for word of mouth.

Step 3: correlate with the timeline

Once you know where the drop is, line it up against a change log: releases, experiment starts and ramps, infrastructure incidents, pricing or packaging changes, marketing starts and stops, and app-store review or ranking changes. The precise shape matters — a cliff on a specific day points at a discrete event, while a gradual slope over the week points at a ramping experiment or a slow-moving external factor.

Check any running A/B tests specifically. An experiment ramped from 5 to 50 percent mid-week can produce exactly this pattern, and it is both a common cause and an easy one to miss because the change is intentional.

Step 4: quantify before you act

Decompose the drop so you know how much of the 15 points each candidate cause explains. If the Android regression accounts for 11 points, that is the story and the rest is noise. If your best explanation accounts for 3 points, you have not found the cause yet and should say so rather than presenting a partial finding as an answer.

What to say when asked for a status

The strong version of the "what would you tell the CEO" follow-up is short and honest: what you know, what you have ruled out, the size of the impact, the current best hypothesis with its confidence, what you are doing next, and when you will report again. Naming what you have ruled out is the part candidates omit, and it is what makes the update useful rather than reassuring.