Metric definitions

The first deliverable in most engagements here is a definition document, not a chart. It names each figure the business runs on, states what that figure counts, states what it excludes, names the system it is read from, and names who settles the answer when it is disputed. The document is short. It is also, reliably, the part of the engagement that produces the longest argument, and the length of that argument is the evidence that writing it down was necessary.

A definition that lives only in a document drifts from the query that implements it, usually within a quarter and usually without anyone noticing. So the derivation travels with the metric in configuration rather than beside it in prose: what it counts, its exclusions, its source system, its refresh cadence. The field is required and carries no default, which means a figure whose derivation nobody can state cannot be added to a reporting surface at all. There is nowhere to put it.

Exclusions are the part that gets skipped and the part that later causes the dispute. Refunds and chargebacks, cancelled orders, internal test transactions, staff purchases, the single outsized account that distorts every average, the trial that converts and then churns inside the same month — each of these is either in or out, and whether it is in or out is a decision somebody has to make. Leaving it undecided does not avoid the decision; it distributes it across every system independently.

Time is the second reliable source of disagreement, and it is more subtle than exclusions because nothing about it looks ambiguous. Last month by transaction date, by settlement date, by invoice date and by the date a subscription period began are four different months. Add timezone and the boundary moves again: systems default to their own, and two reports built from the same underlying events will disagree about the last day of every month for as long as they both exist. The definition names the timestamp and the timezone, because those are properties of the metric rather than of the query.

Grain is the third. What one row represents — an order, a line item, a shipment, a session, a device, a person — determines every average and every count computed on top of it, and it is almost never stated. A conversion rate computed over sessions and a conversion rate computed over people are different measurements that share a name, and a business comparing one against the other is comparing nothing.

Definitions change, and the change is the dangerous event, not the definition itself. When a figure starts counting something slightly different, both the new derivation and the date it took effect are recorded, and the prior series is preserved rather than recomputed backward. A chart that silently restates its own history is worse than no chart, because it destroys the only evidence that anything changed. Anyone comparing this quarter against last is entitled to know whether the two numbers were produced the same way.

Every figure is traceable to the rows that produced it. A number on a surface opens into the records behind it, because a number that cannot be opened is a number that will eventually be doubted with no mechanism for settling the doubt. The first time a figure looks wrong, the difference between an afternoon and a fortnight is whether somebody can click it.

Reconciliation runs on a schedule against the system of record instead of on demand after a disagreement. The billing platform is authoritative for money and the scheduling system is authoritative for appointments; a reported total that has drifted from either is a defect with a date attached, not an anomaly somebody happened to notice. Discovering drift six months late means every decision made in between was made on a number that was already wrong.

Naming carries as much of the definition as it can bear. A label reading active users invites every reader to supply a private meaning; a label reading accounts with a session in the trailing twenty-eight days is unwieldy and correct. Where brevity is genuinely needed for a chart, the full derivation sits one interaction away instead of in a document nobody opens after the week it was written.

The number of defined metrics is deliberately small. A business does not run on forty figures — it runs on somewhere between five and a dozen, and the rest are diagnostics consulted when one of those moves. Defining everything is a way of defining nothing, because a definition list nobody can hold in their head gets consulted with the same frequency as one that does not exist.

Some figures cannot be defined honestly from the data a business currently has, and the output of this work is frequently a list of exactly those. Stating that a number is not presently producible is more useful than producing one whose derivation would not survive being written down, and it converts an open question into a scoped piece of engineering with a cost attached.

None of this is analysis. It is bookkeeping performed on the vocabulary a business already uses about itself, and it is genuinely dull work. It is also the entire reason two reports agree, and there is no version of this that skips the definitions and arrives at agreement later — the disagreement simply stays undiscovered until it surfaces in a room where somebody is being asked to explain it.