KPI definitions in e-commerce: metrics nobody negotiates anymore

There is a meeting that anyone responsible for numbers in e-commerce will recognise. It is scheduled for an hour, the question on the table is whether a measure worked — and the first twenty minutes go on the fact that three people have three different revenue figures for the same month on screen. That leaves forty minutes for the actual decision, and it ends up being made by whoever's export looked most convincing.

In the first article of this series this was one of five patterns that break measurable strategies: every metric has three definitions. This article takes that one pattern apart — because it is the one that blocks all the others. Without a binding KPI definition, no planned trajectory holds up, no impact measurement is verifiable, and no forecast is more than an extrapolation on soft ground.

The good news: this is not an analytics problem and not a question of BI tooling. It is definition work, and it takes days rather than months. The uncomfortable news: it cannot be delegated, because every definition contains a business decision.

Why the name of a metric is not a metric

"Conversion rate" sounds like a defined term. In practice it is a question with at least four open slots: orders divided by sessions or by users? Do orders that were later cancelled count? Do sessions without consent count if they are modelled? And does this apply to the whole shop, or per channel, per device, per language?

Each slot is harmless on its own and decisive in combination. A conversion rate on a user basis sits well above one on a session basis, depending on assortment and return visit behaviour — in Google Analytics a session starts when a user views a page with no active session and ends after 30 minutes of inactivity by default (Google Analytics Help). Someone who researches on Monday and buys on Thursday is one user with two sessions. Both numbers are correct. They are simply not the same number.

The same holds for every other metric that shows up in strategy documents. And as long as the definition is not written down, every meeting negotiates the number first and the decision second. That is not sloppiness, it is a logical consequence: if three departments hold three legitimate views and none of them is marked as binding, the view that happens to fit best is the one that wins.

The anatomy of a KPI definition

A usable KPI definition answers ten points. None of them is optional, and for none of them is a heading enough:

1. Name and question. Which decision is this number supposed to enable? A metric without an attached decision is a report, not a KPI.
2. Formula. Numerator and denominator, written out. Not "revenue per order," but which sum over which line items.
3. Source system and fields. Which system is the binding source — and which fields exactly. Not "from Shopware," but which table, which status, which price basis.
4. Time basis. Order date, invoice date, payment date or shipping date. This choice shifts monthly figures by double-digit percentages whenever delivery times cross a month boundary.
5. Scope. What is excluded: test orders, staff orders, cancellations, B2B or B2C, foreign currencies, marketplace channels, shipping, VAT.
6. Granularity and dimensions. At which levels the metric is valid — total, per channel, per assortment group, per country. Not every metric is meaningful at every level.
7. Target and benchmark. A conversion rate of 2.1% is neither good nor bad without a point of comparison. Without a target it is merely observation.
8. Owner. A person, not a team. They decide on interpretation in case of doubt, and on changes to the definition.
9. Stability rule. From when is the value final? Returns, cancellations and invoicing keep changing a monthly figure for weeks afterwards.
10. Known limits. Where the number is systematically biased, and why. A documented weakness is more reliable than a clean figure whose gaps exist only in one person's head.

In practice this fits on half a page per metric. For the three most important metrics of a goal that is an afternoon's work — and it replaces the twenty minutes of number negotiation in every future meeting.

Written out, such a definition looks like this:

KPI             Net merchandise revenue
Question        How much merchandise value did we keep this month?
Formula         SUM(line item net) − SUM(returns net)
Source          Shopware 6 — order line items, net price basis
Status          paid, partially paid; excluding cancelled
Time basis      Order date (returns assigned to the originating month)
Scope           excluding shipping, excluding test orders (internal customer
                group), foreign currency at the order day's booking rate
Granularity     total, per sales channel, per assortment group
Target 2027     EUR 1.2m · benchmark previous year EUR 980k
Owner           Head of E-Commerce
Stable from     T+45 days (return window + processing)
Limits          marketplace revenue not included — separate metric
Version         v2, effective 2026-10-01 (previously: gross, before returns)

The last point is the one most often missing. A KPI definition gets changed, and that is normal — assortments shift, systems are added, a scope rule turns out to be wrong. But if the change happens silently, nobody can explain the jump in the time series any more. A definition therefore needs a version with a date and a reason, exactly like the guiding statement from part one.

The five metrics most often broken in shops

There are dozens of metrics in e-commerce. Disagreement almost always concerns the same five.

Revenue

The metric with the most variants: gross or net, with or without shipping, before or after returns, with or without cancellations, order date or invoice date, foreign currency at which rate. Every combination has a legitimate use — finance needs the invoice date, marketing needs the order date, purchasing needs the net merchandise volume.

The mistake is not having several variants. The mistake is calling all of them "revenue." If you need three variants, name three metrics: gross order value, net merchandise revenue and invoiced revenue. Technically this hangs off the states of the three state machines that start with every order — order, transaction and delivery (Shopware Developer Docs). Which states feed into the metric is a business decision and belongs in the definition, not in a SQL query only one person knows.

Conversion rate

Settle the denominator — sessions or users — exclude bots and internal traffic, and decide how non-consented traffic is handled. Google Analytics models the behaviour of users who decline analytics cookies based on similar users who accept them (Google Analytics Help). That is useful for trend statements and unusable as the numerator of a revenue-relevant metric.

From this follows a rule that saves a lot of argument: orders come from the shop, sessions come from analytics. The conversion rate is therefore a metric spanning two systems — and for exactly that reason it needs a definition recording which system supplies which part. The same applies to intermediate steps: what counts as a key event is a configuration, not a law of nature (Google Analytics Help).

Contribution margin

The most important and most rarely defined metric. From net merchandise revenue, subtract: cost of goods, shipping cost as actually incurred rather than as a flat rate, payment fees per payment method, returns cost including processing and loss of value, discounts and vouchers. What is not subtracted because it cannot be attributed — warehousing, staff, tooling — is left out deliberately and documented as a limit.

This metric decides almost every prioritisation in assortment and marketing. Without it, you steer growth on a revenue basis and learn about your margin in the annual accounts.

CAC and ROAS

Two questions to settle before any calculation. First: which costs count? Media spend only, or agency fees, tool cost and internal hours as well. Both variants are usable, they just are not called the same thing. Second: which revenue counts? Attributed revenue depends entirely on the chosen attribution model (Google Ads Help), and which model applies is a company decision, not a tool's default.

There is a further point that routinely gets lost in practice: a ROAS on a revenue basis is not a statement about profitability. Google Ads optimises bidding strategies against the conversion value you pass in (Google Ads Help). If that value is the gross order value, the system optimises towards large baskets — including the ones with the highest return rates. Only a conversion value close to contribution margin points the bidding logic at what actually stays in the business. We will cover channels and attribution in detail later in this series.

Return rate

Per item or per order? On a unit or a value basis? And over which time window — because a return arrives weeks after the order. Dividing this month's returns by this month's orders systematically understates the rate; the clean approach assigns returns to the originating month of the order, with a stability rule for the point at which the value counts as final.

The order of magnitude justifies the effort. The Returns Management Research Group at the University of Bamberg expects around 560 million return parcels in Germany in 2026, after roughly 542 million in 2025; close to one in four parcels shipped in German B2C e-commerce is returned (retourenforschung.de). Keeping three variants of a metric that describes about a quarter of your volume is expensive.

One metric, one source of truth

Shop, analytics, ad account and Search Console will never show the same numbers. They measure at different points, in different time zones, with different consent coverage and different attribution. Trying to reconcile that gap away is wasted time.

The productive move is the opposite commitment: every metric has exactly one leading system — the one that sits closest to the event.

• Orders, carts, returns, cancellations → shop
• Sessions, entries, funnel behaviour → analytics
• Cost, clicks, impressions on paid surfaces → ads
• Impressions, clicks and position in organic search → Search Console
• Invoices, incoming payments, credit notes → ERP or finance

Anything computed across several systems gets its composition written into the definition — including the shared key used to join. The interesting metrics arise exactly at those seams: contribution margin per channel, revenue per organic click, return rate per campaign. And so do the misattributions, when the key is not clean.

How many metrics a goal carries

The typical failure mode is not a shortage of metrics but a surplus. A dashboard with forty numbers signals control and delivers none: if everything is watched, nothing is prioritised.

A three-way split works. Every goal carries one lead metric it is measured by. Alongside it sit two or three drivers — the metrics through which the lead metric is moved. And alongside those sit guardrails: metrics that are not meant to improve but must not deteriorate. Conversion optimisation with a return-rate guardrail is a different task from one without. This reduction to a few metrics that relate to each other is the actual substance of the Balanced Scorecard — and it has not worn out in more than three decades.

Everything else is diagnostics. Diagnostic metrics are useful when a lead metric tips over and you want to know why. They belong in the analysis, not in the goal structure.

Four anti-patterns a definition will not fix

A ratio without an absolute. A conversion rate that rises because traffic collapsed is a deterioration. Every ratio needs its numerator and denominator in view.

An average over a skewed distribution. An average order value of EUR 85 can consist of EUR 85 orders, or of very many small ones and a few very large ones. For assortment, shipping logic and free-shipping thresholds those are two entirely different shops. Where the decision depends on the distribution, a median or a percentile belongs in the definition.

A metric without a data basis. Many metrics from strategy documents cannot be computed at all with the systems that are actually connected. Knowing that early is worth more than a nice target figure — because then the first measure is the data integration, not the optimisation.

A target metric without a guardrail. Any metric a team is measured on will be optimised — including by routes nobody intended. Lowering the return rate by making returns harder. Lifting the conversion rate by switching off expensive traffic. Which is why every target metric needs the counter-metric that makes the abuse visible.

How we model this in nuonic Lens

We ran this definition work in spreadsheets and Confluence pages for years, and it decayed every time — not because it was wrong, but because the definition and the number lived in two different places. Whoever sees the number on the dashboard does not go and look up the definition.

In nuonic Lens the KPI is therefore not a field on a dashboard but the object itself: a formula over named fields of the connected sources, with time basis, scope, target, benchmark, owner and stability rule — versioned, with a date and a reason for every change. Every number in the system can be expanded down to the fields it comes from. And where the connected sources cannot produce a requested metric, the system says so instead of filling the gap with an approximation.

nuonic Lens is still in development — there is no public sign-up and no pricing for the time being. If you want to follow the series or talk to us about it early, just get in touch.

The next step

Three checks that need no new tooling:

1. Take your most important metric and have two people calculate it independently. If the results differ, you do not have a metric, you have a term.
2. Write out the ten points above for your three most important metrics. Every point where you have to ask the owner is an open business decision — and that is precisely the source of the discrepancy.
3. Check that every metric can be computed with the systems you have connected. What cannot be computed is not a goal, it is a data project.

Before this series moves into the specialist topics, we cover the fundamentals: the next article is about the goal systems themselves — OKR, Balanced Scorecard, KPI trees and Hoshin Kanri side by side, and which of them fits which shop. Because the definitions from this article need a structure to hang in.

If you would rather not build the goal structure and the KPI definitions alone: that is exactly what we do in consulting, and in growth we run it as a cycle with before-and-after measurement. Get in touch — a conversation with the people who also build it.

Sources and further reading

• Shopware Developer Docs: Orders (order, transaction and delivery state machines)
• Google Analytics Help: About Analytics sessions
• Google Analytics Help: Behavioral modeling for consent mode
• Google Analytics Help: About key events
• Google Ads Help: About attribution models
• Google Ads Help: Target ROAS bidding
• Returns Management Research Group (University of Bamberg): return parcels 2026
• Harvard Business Review: The Balanced Scorecard — Measures that Drive Performance

Keep reading

• Measurable strategy in e-commerce: why goals without numbers aren't goals
• F.O.C.U.S. — the method behind our projects
• Consulting: decisions based on data
• Growth: scaling is a system
• E-commerce search: why your shop needs search that actually finds