Descriptive Analysis

The situation this term answers is one where the records exist and nobody can answer a question about them. Signups and payments land in the database every day, but ask how many people signed up last month and three different answers come back. One person counts who pressed the signup button, another counts who finished email verification, and a third goes from memory.

So meetings stalled in the same spot every time. "The team plan seems to be selling well lately" runs into "I actually felt like it dropped," and neither side has anything but a feeling, so nothing gets settled. Thirty minutes later the meeting ends with someone agreeing to go check the number.

Descriptive analysis is the agreement to settle that part first. Before looking for causes or forecasting anything, pin down what happened as one set of numbers everyone works from. The order is not optional: without knowing what happened you can't look for a cause, and without a cause, forecasts are arithmetic with no meaning attached.

It also sets a quality bar that later work inherits. A cause found on top of a shaky count is a shaky cause, and nobody downstream can tell the difference.

Define what you're counting, pull it for a fixed period, and put it next to something to compare against. Six steps.

1. Write the question and the period in one sentence. "Signups and paid conversions last month" on paper keeps the subject from quietly shifting halfway through.
2. Pin down the definition of what gets counted. Treating an active user as anyone who logged in, versus anyone who opened at least one document, moves the number by a factor of two. Leave the definition unwritten and next month somebody else brings a different figure.
3. Pull the data and look at its condition. Filter out empty rows, the same person counted twice, and test records made from internal accounts. Skip this and thirty staff test accounts end up inside new signups.
4. Summarize. Produce totals and ratios, but don't stop at a single average. If documents per account average 34 while the median is 6, a handful of large accounts are pulling the average up and most accounts actually sit around 6.
5. Put a reference next to it. The figure 4,120 on its own says nothing about whether that is high or low. Lined up against 3,900 the month before and 2,700 in the same month last year, it becomes readable.
6. If it's a number you'll look at repeatedly, put it on a dashboard. Redoing the same calculation by hand every month produces a slightly different answer every month.

What comes out is one table or a few charts. When it reads as a sentence — 4,120 signups last month, 210 of them converting to paid, a 5.1 percent conversion rate — this step is finished. Anything beyond that sentence, including what caused the number, belongs to a later step.

Use it whenever people need to be working from the same numbers. Monthly reviews, the metrics summary that goes to investors, and the operational numbers someone checks every morning all live here. Noticing that yesterday's signups came in at a third of normal is also its job. The signal that something happened arrives here first, even though the reason doesn't.

There are three limits.

1. It can't answer why. That conversion dropped from 12 percent to 8 percent is where this analysis ends; going in to find which step and which group it dropped in belongs to diagnostic analysis.
2. Aggregates hide differences between groups. If an overall conversion rate of 9 percent is really the midpoint between 21 percent on the team plan and 4 percent on the individual plan, no actual user converts at 9 percent. Cutting by step calls for funnel analysis and cutting along time calls for retention analysis.
3. Skewed inputs produce skewed summaries. Average satisfaction collected only from current customers never includes why the people who left, left. That particular skew is survivorship bias.

So two questions are worth asking before carrying the result out of the room. Do the people behind this number represent everyone, and is this number enough on its own to decide from?

"4,120 signups last month, 210 paid conversions. That's 5.1 percent."

A report of what happened, with no cause or recommendation attached.

"This is descriptive only. I haven't looked at why it dropped yet."

The numbers are in; the interpretation hasn't started.

"Give me the median alongside the average. I think a few large accounts are pulling it up."

One summary statistic on its own hides the distribution.

Pagedeck's monthly metrics review opens with this. Mia puts up a single table of signups, paid conversions, active workspaces and cancellations, each against the last twelve months. No interpretation is attached yet, because the point of that slot is getting all eight people in the room looking at the same numbers.

Last review, one cell looked off. Active workspaces were up while documents created were down. Mia didn't guess at a reason on the spot — she wrote down that she'd break it apart the following week. Splitting by plan a week later showed the drop was confined to the individual plan, and from there it was a different kind of analysis.

The most common mistake is reading this as the easy step and skipping past it. Counting things looks like nothing hard, and causes and forecasts look more valuable. But a loose definition here shakes everything built on top of it. Chase a cause without having written down what an active user is, and two months later a different analyst runs the same analysis on a different definition and reaches the opposite conclusion, with no way to settle which one is right.

Assuming the numbers came with their causes attached is the other frequent error. Stare at a table and one cell stands out, and reading it next to the cell beside it produces a story: ad spend went up that month and so did signups, so the ads worked. A competitor may also have raised prices that month, or it may be seasonal. Two values moving together is correlation, and calling one the cause of the other takes separate work. Conclude too early here and next quarter's budget goes to the wrong place.