YouTube watch time dropped 5% last Tuesday: how do you investigate?

In this article

1. Before you start
2. Why interviewers ask this
3. Investigating a watch time drop from first principles
4. Worked example
5. Trade-offs and edge cases
6. How to answer it in an interview
7. Common mistakes
8. Follow-up questions
9. Sources

Before you start

This page is one case in the product and analytics case interviews learning path. It starts from zero.

Why interviewers ask this

A drop in a headline metric is a common alarm in product analytics, and a wrong first guess sends a team chasing a cause that does not exist. Prompts like this one, framed around a video platform's watch time, are reported in product-analytics interviews at large tech companies; no public source confirms how any company asks or scores it, so treat it as a typical open-ended metric case.

The question separates candidates who jump to a theory from candidates who narrow the problem. A weak answer lists causes: "maybe a bug, maybe a competitor, maybe a holiday." A strong answer orders its checks from cheapest and most likely to most expensive, says what each check would show, and stops when one explanation fits. A Staff-level answer also names the traps: a fair baseline, noise, and a mix shift that makes an average move without any behaviour changing. The same shape also answers related prompts, such as rising churn or metrics that move in opposite directions.

Investigating a watch time drop from first principles

Watch time

Watch time is the total time viewers spend watching videos in a period; YouTube's application programming interface (API) reports it as the number of minutes users watched (YouTube Analytics API metrics, as of October 2026). A view that stops early still adds its minutes. It equals the number of views multiplied by the average time of one view, which is why a drop in it can come from fewer views, shorter views, or both.

Segment

A segment is a group of viewers or views that share one property, such as the same device, country, traffic source or content type. Splitting a metric by segment shows which group moved, because a total can hide a large change in one group.

In plain words

Think of a shop owner whose till total is down on a Tuesday. She first checks that the till counted properly. Then she compares with earlier Tuesdays, not with a busy Saturday. Next she asks whether fewer customers came or each spent less, and then which branch fell short. Only after that does she ask why. The analogy breaks in one place: a viewer can produce many views, and a view can be partial, so "customers times spend" maps to "views times view length" only loosely.

In one precise sentence, you investigate a metric drop by making sure the drop is real, splitting the metric into factors, finding the segment where the factors moved, and only then testing causes.

Step 1: Confirm the drop is real and pick a fair baseline

A drop is real only when the data are complete and the comparison is fair. First clarify the metric and scope: which definition of watch time, which region and product, and whether the drop is one day or continuing. Then check the data: did the counting pipeline run fully, did a logging change ship, and does a second source, such as server logs, show the same dip?

A baseline is the usual value you compare against. Use the same weekday, because viewing habits differ across the week. In the worked example, the four earlier Tuesdays average 100.0 million hours (illustrative) and last Tuesday is 95.0, a change of -5.00%. Monday at 92.0 would make last Tuesday look like a rise of +3.26%.¹ The earlier Tuesdays vary by about 0.8 million hours (a sample standard deviation), so a gap of 5.0 is roughly six spreads wide, though four points give only a rough estimate.¹

Input: the daily series. Output: a confirmed drop against a fair baseline.

Step 2: Split watch time into views times average view duration

Watch time equals views multiplied by average view duration, so a drop must come from fewer views, shorter views or both. Average view duration is the total viewing time divided by the number of views. In the worked example, views fell from 1,000 to 970 million (-3.00%) and average view duration fell from 6.000 to 5.876 minutes (-2.06%); 0.97 times 0.9794 is about 0.95.¹

Now the trap. Every platform had exactly the same view length as before. The average fell only because the mix changed: the television app, whose views run long, lost views, so a larger share of the remaining views came from short mobile views.¹ That is a mix shift. A close relative is Simpson's paradox, where an association can reverse when a population is split into subgroups or merged back.

Input: total watch time. Output: which factor moved, and a warning to check the mix.

Step 3: Slice by segment until one slice explains the drop

Slicing by segment finds where the drop lives. Take the dimensions that are cheap and plausible: platform, country, traffic source, content type, and new versus returning viewers. For each slice, divide its change in hours by the total change in hours; a slice that is large and moved a lot is the lead.

In the worked example the television app fell from 50.0 to 45.0 million hours, while mobile stayed at 40.0 and desktop at 10.0, so the television app carries 100% of the 5.0 million hour drop.¹ Its views fell 10.00% while its view length did not change.¹ Then go one level deeper inside the lead, for example by app version, until the slice is small enough to name.

Slicing has a cost. Cut enough ways and some slice will look odd by chance, so decide the dimensions before you look, and treat a small slice with a large percentage change as a hint rather than proof.

Input: the factor that moved. Output: a segment, and a smaller slice inside it.

Step 4: Match the slice to a cause and decide

Knowing where the drop is does not tell you why. Four families of cause cover most cases: a data fault, a calendar effect, a product change, and an outside event. Their signatures are listed below. For the television app, ask whether one app version stopped sending events, whether a release changed the home screen, and whether an outage or a local event overlaps the drop.

Check the release and experiment log first, since a change you shipped is the cheapest cause to test. A data fault means fixing the logging and restating the numbers. A product change means weighing a rollback against what the change was meant to gain. An outside event may call only for monitoring. Finish by naming a monitor, such as an alert on the television app's daily views, and what would change your mind.

Input: the slice. Output: a ranked cause, an action and a monitor.

1 · Confirm and compare2 · Split the metric3 · Slice by segment4 · Match to a causeReplay

Step-by-step description

1. Check the data and use the same weekday.
2. Views times average view duration.
3. Find the segment that moved.
4. Test the cause families and decide.

• The four steps run in order.
• The first step confirms the data and compares the same weekday.
• The last step matches the slice that moved to a cause.

The four steps run in order. The first step confirms the data and compares the same weekday. The last step matches the slice that moved to a cause.

Worked example

The numbers below are illustrative, not YouTube data. Last Tuesday's watch time is compared with the average of the four earlier Tuesdays, split by platform. Each platform's hours equal its views times its average minutes per view, divided by sixty.¹

The television app held 30.00% of views and 27.84% of them last Tuesday.¹ Weighting last Tuesday's segment lengths by the old mix gives 6.000 minutes, so the whole 2.06% fall in average view duration is a mix shift. A reader can redo any row with a calculator from the views and minutes in the saved script.¹

Text description of the diagram

• Television app watch time fell from 50.0 to 45.0 million hours.
• Mobile stayed at 40.0 and desktop at 10.0 million hours.

Television app watch time fell from 50.0 to 45.0 million hours. Mobile stayed at 40.0 and desktop at 10.0 million hours.¹

Trade-offs and edge cases

Four families of cause leave different signatures, so the first check differs for each. A data fault shows as one source disagreeing with another. A calendar effect, such as a public holiday, a school break or a clock change, shows as a dip that lines up with a known date and with the same period last year. A product change starts at a release time in one segment. An outside event, such as an outage or a major news or sports event, appears across many segments in one region.

Text description of the diagram

• A data fault shows as one source disagreeing with another.
• A calendar effect shows as a dip that lines up with a known date and with the same period last year.
• A product change starts at a release time in one segment.
• An outside event appears across many segments in one region.

Edge cases follow from the method:

• Several causes at once. One slice can hold a data fault and a real change together. Quantify each part before you call the case closed.
• Too many slices. Slicing many ways finds chance patterns. Fix the dimensions first and confirm a lead on a second period.
• Reversals. The Stanford Encyclopedia of Philosophy explains that an association in the whole population can reverse inside each subgroup, and vice versa. A total and its parts can tell opposite stories, so state which level you are reasoning at.
• Restated data. A number that changes when you rerun it later means the first reading was incomplete; wait, then redo the baseline.

For more worked prompts of this kind, browse the interview guides in the learn library.

How to answer it in an interview

The spoken structure: say it is real, say the baseline, split the metric, slice, then name the cause family and the action. Practise it aloud once; a 45 to 60 second answer leaves room for the interviewer's follow-ups.

Common mistakes

• "Watch time fell, so the recommendations broke." (wrong) This picks a cause before checking anything. (right) "First I confirm the data and the baseline, then I look at which part of the metric moved."
• "It is higher than Monday, so nothing is wrong." (wrong) Weekdays differ in viewing habits. (right) "I compare with the same weekday in earlier weeks."
• "Average view duration fell, so viewers are losing interest." (wrong) A mix shift can move the average with no change in behaviour. (right) "I check whether the mix of views moved before reading the average as behaviour."
• "The television app dropped, so that is the cause." (wrong) A slice shows where, not why. (right) "That is where the drop lives; the release log and the second source tell me why."

Follow-up questions

Churn has increased over the last two months. What do you do?

The shape is the same. Clarify which churn is meant (customers lost or revenue lost), confirm the data and a fair baseline, then split the rate into who joined and who left and segment new against existing users before testing causes.

What do you do if your primary metric goes up but engagement duration goes down?

Split both with the same factor logic and look for a mix shift, for example many new users with short sessions. Then decide which metric reflects the goal and keep the other as a guardrail. See also the data science interview questions for neighbouring prompts.

Metrics moved in different directions: how do you interpret the results and decide next steps?

Check each metric's definition and segment each one, because a total and its parts can disagree. Pick the metric that matches the product goal, state the trade-off in words, and propose the next test. The ads and marketplace domain questions show the same reasoning with two-sided metrics.

How do you know whether a five percent drop is just noise?

Compare the drop with the spread of earlier values for the same weekday: the further outside it falls, the less likely it is chance. Use more than four weeks where you can, since a few points give a rough spread.

What if the drop were only one percent, or only on weekends?

A smaller drop sits closer to normal spread, so noise becomes a live explanation and you would wait for more data. A weekend-only drop is a segment finding: compare it with earlier weekends, then check for a release schedule, a known date or an outside event that touches weekends. The Meta product architecture interview guide covers a nearby product round.

Sources

1. Simpson's Paradox (Stanford Encyclopedia of Philosophy) — Stanford Encyclopedia of Philosophy, 1 Apr 2016
2. YouTube Analytics API: channel and content owner metrics — Google for Developers (page last updated 2026-09-14 UTC), 14 Sep 2026
3. NIST/SEMATECH e-Handbook of Statistical Methods, 1.3.5.6 Measures of Scale (live page, undated; accessed 2026-10-26) — NIST, 26 Oct 2026