40+ KPIs, four dashboards, and one agreed definition each

There is a particular kind of meeting that happens in every large retailer. Two teams present sales figures for the same week and the numbers do not match. Everyone spends forty minutes discovering that one team counted orders at checkout and the other at fulfilment, and that both were right.

We built four dashboard families for a retailer's analytics platform. The engineering was substantial. The part that determined whether anyone trusted the output was upstream of it: writing down what every metric meant, and getting the business to agree, before building anything.

Digital: are we ahead of plan, right now?

The anchor of the digital dashboards was demand versus plan — a line chart of actual sales against a planned figure derived from last year's data, extrapolated forward.

The subtlety is time grain. A plan is not one number; it is a different number at every zoom level. The daily view needed hourly grain, the weekly view daily grain, and so on up through monthly, quarterly and annual. As the ingestion framework brought sales metrics in, each was plotted against the plan at the matching grain, so "are we ahead?" had an answer at whatever resolution the person asking cared about.

Around it sat the metrics that explain movement rather than just report it:

• Device type — desktop, mobile web, native app, tablet, kiosk. Not trivia: it tells you where to invest when one channel is quietly carrying growth.
• Marketing channel — affiliate, email, organic search, direct, paid search.
• Credit events — a flag for whether a promotion, cashback or discount was running. This one is easy to omit and expensive to omit, because without it "sales are up on last year" is unanswerable. Up because of what?
• Visits, units, conversion rate, orders, plus average unit retail and average order value, each trended against last year.

The conversion funnel was the piece people used most: visits → visits with product view → visits with cart additions → visits with checkout → orders. A funnel is not a report, it is a diagnostic. A sharp drop between cart and checkout is a question — price, a technical fault, or simple abandonment — and it feeds directly into cart and browse abandonment campaigns.

Stores: the same questions, a different shape

Physical retail needs its own vocabulary. Comparable sales, transactions, returns value, average transaction value, units per transaction.

Two design choices are worth pulling out.

Territory versus climate. The store dashboard could be sliced by territory or by climate band — hot, cold, mild, very hot. For a business selling seasonal goods that second axis explains more than geography does. A warm January in two distant regions is a better predictor of what sold than which sales district each store belongs to.

A state-level heatmap. Store demand sales rendered as a choropleth, darker meaning higher. It answers "where are we strong and weak" faster than any table, and it is the view an executive actually opens.

Payments: the questions nobody thinks to ask

The payment dashboards look mundane and repeatedly were not. Revenue split by payment type — the major card networks, gift cards, merchandise return credit, the store's own credit card — and by checkout type: normal shipping, express shipping, and the various wallets.

Wallet adoption got its own view: how many customers have one versus not. That is a leading indicator. Wallet users check out faster and abandon less, so the adoption curve predicts conversion changes before conversion moves.

Pathing: what people do instead of buying

The pathing dashboard counted where sessions actually went — exited site, search, cart, sign-in, store locator, shop menu, navigation shortcuts, and into specific departments.

The most useful bar on that chart is exited site. Everything else tells you what engaged people did; that one tells you how many did nothing, which is usually the largest number and the one least discussed.

The real lesson

The dashboards were the easy part. Ingesting the data, computing metrics at multiple grains, and rendering it in near real time is well-understood engineering.

What made the project work was the phase before it: writing down every metric definition and having the business agree. Over forty KPIs, each with a statement of what it shows and why the business needs it — documented, not assumed.

Skipping that step does not save time. It defers the argument to the moment two dashboards disagree in front of an executive, and then the argument costs far more, because by then it is also a credibility problem.

If we were starting this again, the order would be identical: agree the definitions, then build the pipeline, then draw the charts. The chart is the last and least of it.

On numbers

The source documentation records typical values for some metrics. Those are the client's commercial figures and are not reproduced here. The structural detail above — the dashboard families, the metrics, the time grains, the funnel stages, the territory/climate split — is drawn from the project's own specification.