Drill-down

Screens that gather numbers together have been around for a long time. The trouble was that such a screen produced questions rather than answers.

A dashboard collects key metrics in one view, and almost everything it displays is a total. The line "2,000 signups this week" carries one fact — a thousand fewer than last week — and withholds every useful detail. Whether ad traffic dropped, whether search traffic dropped, whether one country accounts for all of it: none of that fits in the line.

So everyone who read the screen stalled in the same place. At Pagedeck, Mia would see the Monday number bend downward and ask the data analyst to look into why, and it took two days for an answer to come back from a hand-written query. Nobody moved during those two days, and the answer usually triggered a follow-up — "can we see that one channel on its own?" — that cost another day.

Drill-down moves that round trip into the screen. The paths for splitting a total are cut ahead of time, next to the total itself, so the person who got curious follows them right there. Two days between question and answer becomes two clicks, and the person who had the question is the one who gets the answer.

Pick a total, choose one attribute to split it by, then take the slice that stands out and split again. The whole technique is those three moves on repeat.

1. Choose a starting point — the one value on the screen that looks wrong. Signups falling from 3,000 to 2,000.
2. Choose an attribute to split by. Acquisition channel, plan, country, device: anything the data can be cut along. Split by channel and search and referral hold steady while ad traffic alone drops from 1,400 to 500.
3. Take the slice that moved and go one level deeper. Split ad traffic by campaign and one of five campaigns sits near zero. At that point there is exactly one place to go check.
4. Decide where to stop. Every split leaves fewer people in each slice, and past some depth the gaps between slices are the size that chance alone produces. A five-person gap in a forty-person slice means nothing, and the fact that it's displayed to two decimal places doesn't make it mean more.

What you can split by is decided by the data, not the screen. An event property is the extra information stored alongside a recorded action, and if signups were never recorded with the acquisition channel attached, no amount of screen design will let you split by channel later. Shallow drill-downs usually trace back to sparse properties at recording time rather than to a badly built dashboard.

Order matters too. Splitting by country first and device second shows you a different picture than device first and country second, because each choice decides what the next level is even able to reveal. Starting with the attribute that divides the largest chunk keeps the number of wrong turns down, and it also keeps each slice big enough to survive another split.

The most common use is narrowing the cause on a day when a metric moves. Conversion drops, so you split by country and check whether one market fell or every market fell a little. One market points at local payment methods or something happening outside the product; an even fall points at a release or an outage. The two paths call for different people, so the split decides who gets pulled in next.

It also pairs with funnel analysis, which measures how many people fall out at each step between signup and payment. When one step shows unusual drop-off, that step gets split again by device or channel. A 60 percent drop-off at checkout that turns out to be 80 percent on mobile is a story about typing on a small screen, not about the checkout screen as a whole — and the fix that follows is a different piece of work assigned to a different person.

A third case follows anomaly detection, which watches for values outside the usual range and raises an alert automatically. The alert tells you something is off and stops there. Finding where it's off is the drill-down's job.

"Can I drill down on this number? I want to see it split by channel."

Asking whether a path from the total into the detail actually exists on this screen.

"Overall we're down 5 percent, but drilling down, it's only Android."

The drop isn't spread evenly — it's concentrated in one slice.

"Two levels down the sample is 40 people. This gap looks like noise to me."

Splitting any further stops distinguishing a real difference from random variation, so this is the floor.

At Pagedeck the move gets used most in the weekly metrics meeting. Mia puts the dashboard up, clicks into whichever line looks wrong, and three or four levels later the affected slice is clear enough to assign on the spot. Anything on the ad side goes to Ravi; anything in the payment flow goes to Theo. Nobody leaves the room with a research task, because the narrowing already happened while everyone watched.

Speed isn't the only thing that changed. Questions raised in the meeting used to turn into analysis requests answered the following week, and in the meantime people argued from their own guesses. Now the guesses get erased in front of everyone, on the same screen.

The most common mistake is expecting a drill-down to hand over a cause. It feels that way because the final view is specific and because digging further always seems possible. What the technique narrows, though, is location rather than cause. "Signups fell inside one ad campaign" is where it ends, and whether that campaign ran out of budget, swapped creative, or pointed at a broken landing page has to be checked outside the screen. Stopping there and calling it an answer leads to fixing the wrong thing, usually something adjacent to the real problem that looked plausible on the last screen.

The other one is believing that deeper means more accurate. It runs the other way. Each split divides the people in a slice by some factor, and three levels down the slices hold a few dozen people each. At that size the differences between slices come out differently every time you look. Summarize such a number as "women in their thirties on Android are having trouble" and next week the same procedure will surface an entirely different slice.