The Ecommerce KPI Framework: Dashboards That Drive Decisions

Download the workbook: an Excel framework, pre-filled with a fictional retailer so every sheet shows something working. No form, no email, no gate.

Most ecommerce dashboards are accurate and useless at the same time. They report forty numbers, every one of them correct, and the meeting spends its hour reading them aloud. Nobody agreed in advance what a movement would mean, who would act on it, or what they would do.

This framework starts from one rule. Every number on a dashboard should name the decision it informs and the person who owns that decision. If it informs no decision, it does not belong on the dashboard. It can live in a report, where nobody has to pretend it matters.

The discipline

What? So what? Now what?

Every KPI in the library, and every line on every dashboard, answers three questions.

What?

What the number is and how it moved. On the dashboards it writes itself from the figures: “−6.3% against plan, +4.0% against last year, worsening”.

So what?

Why that movement matters commercially, in money where possible. Written by the owner before the meeting.

Now what?

The action, the owner and the date. A number that cannot answer this question is a report, not a KPI.

The workbook enforces it. A check fails any decision measure or trigger without a written So what and Now what, and flags any measure at Act with nothing against it. The meeting decides rather than reads.

The library

Four classes, not one list

Every KPI is classed by what it is good for, because a number that diagnoses a problem is not the same as a number that decides what to do about it.

Class	What it means	In the library
Decision	Someone can act on it directly: it has an owner, a threshold and an action	16
Investigation trigger	Movement means dig. It diagnoses; it rarely decides on its own	22
Context	Worth knowing, rarely worth a meeting: static, slow or unconnected to a decision	6
Misleading	Drives the wrong behaviour when used as a target. Replace it	5

Each entry carries its definition, formula, data source, owner, cadence, the movement to watch, the So what, the Now what, how it misleads, and the measure to read it with. Definitions are written tightly on purpose: two teams should not be able to report the same KPI two different ways. Ten spare rows take your own.

The misleading five

Three of them will surprise nobody: email list size, time on site and app downloads. Two will start an argument.

Conversion rate across all sessions. It counts every visit to find a store, read the FAQs or check an order as a failure, and it moves with traffic mix and campaign objective. A visitor who comes three times and buys once is not a failure. Used as a target, it rewards turning away the customers who were not ready to buy. Segmented to the committed buying journey, sessions that viewed a product, it becomes a useful trigger.

From the field. At one business, marketing was targeted on traffic. It bought cheap display, hit the target, and conversion fell; the display traffic bounced 80% of the time. The way I explained it to the CEO: marketing had brought five coaches of customers to the car park outside the window, and opened the doors on one. The CEO still looked at conversion rate afterwards, but asked for the qualification first.

Return on ad spend. It divides revenue the platform attributes to itself by the platform’s own spend. It ignores margin, returns and whether the sale would have happened anyway. Contribution after marketing, and the marginal cost of the last new customer, answer the question it pretends to.

Net promoter score sits in Context rather than Misleading: useful as a long-run signal, too slow and too averaged to act on in a weekly meeting.

Start here

The meaningful test

List every metric on your current dashboard and answer six questions for each.

A named person owns it?
It informs a specific decision?
There is an agreed action when it trips? The Now what.
It moves enough to act on at its cadence?
It is tied to money, margin or customers, not volume?
It is read with a counter-metric?

The verdict is one of five: keep as a decision KPI, keep below the line as a trigger, demote to context, delete, or replace it, which applies whatever the answers if the library classes it Misleading.

The workbook’s sample is a typical dashboard of a fictional retailer, before the test:

Metrics on the dashboard

That drive a decision

To delete or replace

Run it on your own board pack before reading anything else in the workbook. It takes about an hour, and it is usually the hour that changes the next meeting.

The templates

Three dashboards, three cadences

Board: eight measures or fewer, monthly or quarterly. Each names its owner and the decision it asks of the board. Escalate only what needs board resources, oversight or a strategic decision; anything management is already authorised to fix stays below. A red measure with no decision written against it is flagged, because it is a report, not a board paper.

SLT: the weekly levers. Each has a tolerance, a status against plan, and a five-week direction. A measure can be on plan and worsening, and that is the one to talk about. A measure can be off plan and improving, and that one may need patience rather than a new initiative.

Departments: the daily levers by team. Trading, performance marketing, CRM, site and UX, operations and service, each with a threshold, an owner and the first place to look when it trips.

Diagnosis

The driver tree and the investigation paths

When a board number misses, the driver tree splits the variance exactly: net sales into traffic, conversion, basket and returns; contribution into volume, margin rate, cost per order and media. Each piece goes to its owner. In the tree, conversion is arithmetic, not a target.

The tree then reasons in four steps, and keeps them apart:

What the numbers establish. In the sample, traffic beat plan by 4.5%, conversion fell 0.18 points, and conversion accounts for £161,000 of the shortfall.
What the pattern suggests. With media 9.6% over plan, one hypothesis is that bought traffic was less ready to buy. It is a hypothesis, not a finding.
What would establish the cause. Conversion by channel and campaign, and availability on the lines the traffic landed on.
Now what. The largest drag, its owner, and the investigation path to run.

There are seven investigation paths, each five questions long and asked in order: conversion falls, AOV falls, returns rise, acquisition cost rises, availability falls, checkout completion falls, and contribution falls while sales rise. Each question names the KPIs to pull and what it rules in or out, so the team stops arguing about the cause and starts checking it.

Governance

Setting thresholds

Every threshold in the workbook is an example, not an industry norm. Set your own in four steps:

Baseline from reliable history, peak and non-peak separately.
Set the line where a move becomes commercially material and the team can act on it.
Name who may change it, and on what evidence.
Review it on a fixed date, and never move it after the fact to make a bad month look acceptable.

The loop

With the Ecommerce Budget Builder

The Ecommerce Budget Builder sets the assumptions: traffic, conversion, basket, returns, margin and media. This framework watches them. When a measure moves off plan, the driver tree says which assumption broke, the investigation path finds out why, and the reforecast log records the evidence, the decision and the forecast impact in pounds. The budget then moves on evidence, rather than waiting for the plan to recover by itself. Contribution, returns rate and AOV are defined the same way in both workbooks.

Limits

What it does not do

It is a starting point, not a standard. A subscription business, a marketplace seller and a furniture retailer will reclassify some of it, and should.
It does not fix data. A driver tree is only as good as the session, order, returns and margin definitions underneath it.
The bridges measure size, not cause. They change one driver at a time in a fixed order, so the split depends on the order. The investigation paths are there for the why.

Sources

Where this comes from

The classes, the paths and the misleading five come from running ecommerce P&Ls since 1999, and from sitting through a great many meetings where the dashboard was read aloud and nothing changed. The figures in the workbook belong to a fictional retailer. The workbook has been tested against 24 deliberate ways of breaking it, from measures sitting exactly on their tolerance to a misleading KPI creeping back onto the board dashboard.

If the meaningful test removes more than half of your board pack, that is worth a conversation before the next board meeting.

Back to Thinking