Writing
Part 1: Going One Layer Deeper - Why Granular Metrics Matter
Aggregates are comforting because they make the world fit on one chart. They are also where problems go to hide. The signal you need is usually one layer deeper — by cohort, surface, region, device, creative, user state — and the habit this series is about is simple to state: refuse to stop at the top-line number.
I learned this early. At Quora, my first job out of school, the search traffic chart looked fine — traffic was growing. One layer down, traffic per piece of content was quietly declining. The aggregate said healthy; the ratio underneath said something structural was wrong. Pulling on that thread turned into the defining project of my time there, and it permanently changed how I read a dashboard.
Averages give you generic fixes
Product and systems decisions are local. A conversion drop is never just "conversion dropped" — it dropped for someone, somewhere, under some condition. Where is it down? For whom? Which slice moved and which stayed flat? Averages lead to generic fixes, and generic fixes have a way of changing the parts of the system that weren't broken while leaving the broken part alone.
Mixture effects are the classic trap. The top line moves not because any segment changed behavior, but because the mix of segments shifted. Check for this before accepting any trend explanation, or you'll spend a week explaining a change that didn't happen.
Start from the decision, not the data
The mental model I use: design analytics backwards from the decision tree. If this metric moves, what would you actually do? A segmentation earns its place when it maps to an intervention — a ranking slot you can change, a latency bucket you can fix, a policy category you can review. Detail for its own sake just produces dashboards nobody opens twice. The goal is to make the bottleneck visible at the level where action is possible.
Granularity is only as good as the plumbing
One honest caveat. Granular metrics only work when the data underneath holds its shape: stable event names, versioned schemas, identifiers that survive the trip from impression to click to conversion, sampling and backfill rules written down before launch. None of this is fun. All of it is load-bearing. A beautiful slicing tool on top of ambiguous events is just a faster way to be confidently wrong.
The first layer tells you something changed. The second layer tells you what to do. The rest of this series is about building the habit — and the plumbing — to get there.