Why reporting fails
Reporting usually fails for one of two reasons. Either nobody agreed what the numbers mean, so every report reopens the same argument about definitions, or the report is a wall of data with no recommendation attached, so it gets skimmed and forgotten. A data dictionary fixes the first problem. How a report is written fixes the second, covered in report writing concepts. The dictionary is worth building even for a small brand with a handful of channels. The argument it prevents (“is that revenue net of refunds?”) happens whether the brand is small or large; a dictionary just means it happens once, in a session, instead of in every meeting from now on.KPI tiers
A KPI is a metric that changes what you do
Plenty of numbers are worth watching. Far fewer are worth calling a KPI. The test is simple: if this metric moved 20% in either direction, would you do anything differently? If the honest answer is no, it is context, not a KPI. Keep it in the dictionary with the KPI tier Tracked, or leave it off entirely. Starting from decisions rather than from whatever a dashboard happens to surface is what keeps this list short. Starting from available data is how a team ends up reporting impressions and time on page: numbers that are easy to pull and hard to act on.Three tiers explain each other
A flat list of ten KPIs gives no sense of priority or cause. Splitting them into tiers turns the list into a chain, so a report can explain why something moved instead of only stating that it did.
Each tier explains the tier above it. If contribution margin (tier 1) falls, tier 2 says why: efficiency dropped, or order value fell, or acquisition cost rose. If acquisition cost rose, a Tracked metric says why that happened: cost per click went up, or conversion rate fell. A report built this way reads as a chain of cause and effect rather than a list of unrelated facts.
Businesses with a longer path from marketing touch to revenue run into a specific version of this problem. A lead generated this week might not close for months, so the tier 1 outcome (revenue, closed deals) always lags the tier 2 and Tracked activity that produced it. Reporting those lagging outcomes against the cohort that generated them, rather than the calendar period they landed in, is what keeps this month’s spend from getting credited with a deal that pipeline built two quarters ago.
A target needs a basis, and a floor
Write down where each target came from: last year’s number, a board plan, a benchmark, or, honestly, a guess. There is nothing wrong with starting from a guess, but a guessed target quietly treated as a firm commitment is how reporting turns into theatre, where every review becomes a defense of a number nobody actually chose carefully. A floor is different from a target. It is the level that triggers an actual phone call rather than a line in a report. Setting it in advance means the response to a bad number is decided ahead of the moment it happens, rather than negotiated under pressure.Choosing KPIs for your work
One primary KPI per piece of work
Every campaign, experiment and report has to name what it is judged on. Two primaries is no primary. Work judged on two numbers gets reported against whichever one moved favorably, which is how everything succeeds and nothing improves. Name one primary, then name the counter-metric that catches damage done in pursuit of it. Both go in the brief before launch. E.g. A Doughnut Labs trial-acquisition campaign is judged on trial starts, with trial-to-paid rate as the counter-metric. Starts can always be bought by loosening the offer; the counter-metric stops that reading as a win.Which KPI applies to what
Match the metric to the cadence
A metric that cannot move inside the reporting period cannot be reported on usefully. This is the most common mismatch in reporting.What disqualifies a metric
A metric belongs in the dictionary. It does not automatically belong as a KPI. Leave it as Tracked if any of these are true:- A 20% move in either direction would not change what you do.
- Nobody owns it.
- No one can state its definition without looking it up twice.
- It is a rate whose denominator nobody agrees on.
- It only ever moves up, which usually means it counts activity rather than measuring a result.
- It moves so closely with another KPI that the two always tell the same story.
Adding a KPI
- Name the decision it changes. If nobody can say which decision this metric would change, and at what threshold, it is a Tracked metric. Tracked metrics belong in the dictionary and on dashboards; they don’t need a target.
- Define it in the dictionary first. Formula, grain, source and owner. A KPI adopted before it is defined gets defined differently by each person who reports it.
- Name what it replaces. The list is capped. Adding without removing is how 8 useful KPIs become 20 that nobody reads.
- Backfill before you adopt it. A KPI with no history has no normal range, so nobody can tell whether this week’s figure is good. Backfill at least a full seasonal cycle where the data allows.
- Record the change in the decisions and changes section of the dictionary, with the date. A trend line that shifts because the metric changed, with no record of the change, reads as performance.
Retiring a KPI
Keep the definition in the dictionary and set its KPI tier back to Tracked, or mark it retired, with the date and what replaced it. Reports covering earlier periods still reference it, and deleting the definition makes those reports unreadable.Reviewing the KPI set
Review the KPI set once a year, alongside the dictionary. Mid-year changes should be rare: a KPI set that changes often cannot show a trend, which is most of what it is for.The north star metric
It is the tiebreaker, not the most important KPI
A team can have five well-defined KPIs and still argue constantly, because KPIs conflict with each other in ordinary ways: volume up and margin down, lead count up and lead quality down. A north star is the one metric that settles which side of that tradeoff wins when both options look reasonable. There can be twelve KPIs. There is exactly one north star.Four tests, and all four have to pass
Does it reflect real value? Moving it up should mean the business is genuinely better off, not just busier. Revenue passes this test; a metric like sessions does not, because sessions can be bought without buying anything real. Can the team actually move it? A metric that only responds to forces outside marketing’s control cannot guide a weekly decision, however important it is. Does it respond inside the decision cycle? A metric that takes two quarters to shift cannot steer a decision made weekly. Something can be a genuinely important number and still be the wrong choice for a north star, simply because it moves too slowly to guide the cadence of decisions being made. Can it be calculated the same way every time? If two people compute the same metric differently and get different answers, it cannot function as a tiebreaker, because the tiebreak itself becomes a dispute. This is the reason a north star always needs a full formula entry in the dictionary before anyone adopts it: the definition has to be settled before the metric can settle anything else.Every north star needs a counter-metric
Any single metric, optimized hard enough, can be gamed, often without anyone intending to game it. Cut spend and margin looks better while the business quietly shrinks. Lower the bar for what counts as a lead and lead volume climbs while pipeline quality erodes. The counter-metric is what surfaces that tradeoff in the same report, rather than in a postmortem after the damage is visible. A counter-metric does not need its own target. It needs to be reported in the same place as the north star, every time, so the tradeoff is part of the routine reading rather than a special investigation.Attribution
Every model is wrong somewhere, so write down where
Attribution decides which marketing touch gets credit for a result. Every attribution model is wrong in a different way, because no model can see everything that actually influenced a decision to buy or convert. The goal is not finding the one true model. It is knowing where each number comes from and where it is known to be wrong, so that a change in the number reflects a change in performance and not a change in how credit was counted. This is also why platform-reported numbers can never simply be added together. Each ad platform counts a conversion if it touched that customer’s path anywhere within its own window, so two or three platforms often claim the same conversion. Summing them produces a total larger than the number of actual orders, not because anyone is lying, but because every platform is answering a slightly different question.What changes the right choice
Three questions tend to decide which model fits: how long the path from first touch to purchase usually runs, how much of the budget goes to channels that cannot be clicked (audio, connected TV, out-of-home, sponsorships), and how quickly the team needs an answer to act on. A short path with mostly clickable channels can lean on a simple model read weekly. A long path with meaningful unclickable spend needs either a wider attribution window or a different kind of measurement, like a holdout test, alongside it. Attribution sophistication tends to track a business’s maturity, not because more complexity is inherently better, but because more complex models need more history and more volume to trust. A brand just getting started is usually better served by one simple, consistently applied model than by an ambitious multi-model setup nobody has the data to validate yet.Windows differ by browser and platform
An attribution window is how long after a touch a conversion still gets credited to it. Each ad platform sets its own defaults, and browsers limit how long tracking cookies last, so the same conversion can fall inside one window and outside another. No single window holds across all your traffic, which is why the data dictionary does not fix one. Handle it case by case instead. Label every platform-reported number with where it came from, keep a blended check such as total revenue against total spend alongside it, and when a decision depends on the answer, test it. A window that is much shorter than the real path to purchase quietly starves the top of the funnel of credit, and a holdout test is what shows you that.Validate the model, don’t just trust it
An attribution model is an assumption about how the business works, not a direct measurement of it, and assumptions drift out of date. A holdout test, where a channel is turned off for part of the audience and the gap against a matched group is measured, is the most direct way to check whether a model’s credit roughly matches reality. A blended sanity check, comparing total attributed conversions against actual orders, is a lighter, ongoing version of the same idea. Either way, the result is worth recording even when it simply confirms the model still holds, since that confirmation is itself useful information six months later.Source of truth
One approved system per metric
When two systems disagree on a number, the conversation should be “which one is the approved source,” not a debate about which tool is more trustworthy in general. Naming the approved source ahead of time turns that moment from a negotiation into a lookup. A useful test for whether a definition is actually complete: two people pulling the same metric from the approved source, on the same day, should get the same number. If they don’t, the gap usually isn’t in the tool. It’s in the definition, which is missing a rule about a filter, a date range, or an inclusion decision that both people are silently answering differently.Assign by who owns the event
The system that owns a given event is usually the right source for the metrics that event produces: an ecommerce platform for orders and refunds, an ad platform for spend, a CRM for deals, a lifecycle tool for sends. The most common exception people get wrong is treating an analytics platform as the source for revenue. Analytics tools miss transactions blocked by ad blockers or lost to consent rejection, and they don’t know when a refund happens after the fact. Revenue belongs to whichever system actually took the money.Metrics that span systems need a join rule, not just a source
A metric like contribution margin pulls from several systems at once: revenue from one, cost of goods from another, ad spend from a third. Recording the join key and the timing for each input matters as much as recording the systems themselves, because a margin figure that quietly combines today’s revenue with last month’s cost of goods will disagree with finance’s version in a way that looks like an error even when it isn’t one.Ownership
Every metric needs one name attached, not a team. A metric with shared ownership tends to drift silently: someone adds a filter, someone else changes a date range, and months later two views of the same number disagree with no record of why. A single owner does not mean one person does all the work. It means there is one person to ask when the definition needs a decision. Four roles cover most of what a metric needs, and one person can hold several of them:
A report with no decision owner tends to become a status update that nobody is accountable for acting on.
Definitions change, but never silently
A definition should only change when its owner approves the change, the data owner confirms it can actually be implemented, the change is logged with an effective date, and every report owner knows before the next report goes out. A definition that changes quietly is worse than one that never gets fixed, because it makes last quarter’s numbers stop matching this quarter’s without anyone knowing why. When a definition does change, either restate history under the new definition or mark the break clearly on any chart that spans it. A trend line crossing a definition change without a marker is misleading even though every individual number in it is correct.Related resources
- KPI definitions Where you fill in your north star, tier 1 and tier 2 metrics.
- Data dictionary The fill-in workbook for every concept on this page.
- Report writing concepts How the metrics this page defines turn into a report people actually read.
- Data quality and rollup concepts How daily rows roll up into weekly, monthly, quarterly, and annual numbers without breaking.
MCP wiring
MCP wiring
Type: concept
Domains: reporting, strategy
Provides:
Uses:
Domains: reporting, strategy
Provides:
concept:reporting, concept:kpiUses:
article-brief, campaign-brief, data-dictionary, experiment-brief, kpi-definitions, landing-page-brief, report-templateGet the complete Marketing OS
Subscribe to Snow Doughnut, our free weekly newsletter, and get the Notion workspace where you do the work: tasks, a marketing calendar, name generators, briefs, and more.
Have us run it for you
We implement this system, use it to help you grow and scale, then hand it all off to you when you’re ready.