Problem
The tile says “Active users: 41,208”. Marketing means anyone who opened the app this month. Engineering means anyone who made an authenticated request in the last 24 hours. Finance means anyone with a paid seat. All three read the same tile and leave the meeting agreeing on a number none of them share.
Solution
Put the definition next to the number. An info affordance on the panel opening a short, plain-language statement of what is counted, over what window, with what exclusions, and who owns it.
The content is the hard part, and it is four things rather than one.
What is counted. In the viewer’s words, not the column name. “Anyone who made at least one authenticated request” beats count(distinct user_id).
Over what window. Almost every disagreement about a metric turns out to be a disagreement about the window.
What is excluded. Internal traffic, bots, test accounts, refunds, cancelled orders. Exclusions are where two honest implementations diverge and neither party knows it.
Who owns it. A name. A definition with no owner is a definition nobody can correct, and it will be wrong within a year.
The failure mode this pattern addresses is specific and expensive: not that the number is wrong, but that it is right under a definition the reader doesn’t hold. That failure produces confident, articulate disagreement, and it survives any amount of chart polish.
The structural version of the answer is a modelling or semantic layer, where a metric is defined once and every surface reads that definition. That is better than tooltips and much harder to retrofit. The tooltip is what you can do this quarter; the model is what stops the problem recurring.
Use when
The audience is mixed, the metric name is ambiguous, or the number gets quoted outside the room it was shown in. Any tile that ends up in a board deck.
Don’t use when
The metric is genuinely self-evident to everyone who sees it—a CPU percentage on an engineering dashboard needs no gloss, and adding one dilutes the affordance where it matters.
Trade-offs
Definitions rot faster than charts, because the query changes and the tooltip doesn’t, and a stale definition is worse than none: it is authoritative and wrong. Writing them well is slow, and writing them badly produces the sentence “Active users is the number of active users”. Tooltips also hide the definition behind an interaction most viewers never perform, so the people most likely to misread the number are the least likely to check. And a definition that admits a metric’s messiness can undermine confidence in a number that was, in fact, fine.
Checklist
- Does the definition say what is counted in the reader’s language?
- Does it state the window?
- Does it list exclusions—internal traffic, bots, test accounts, refunds?
- Does it name an owner?
- Is the definition stored with the metric, or retyped per dashboard?
- When the query changes, what makes the definition change?
- Is the affordance visible enough that someone who doesn’t know they’re confused might click it?
- Is the same metric defined identically everywhere it appears?
- Does the definition distinguish this metric from the similarly-named one next to it?
- Could this be a modelling-layer definition rather than a tooltip?
Compare
Looker answers this structurally with LookML: metrics are defined in a modelled layer and every chart inherits that definition, so the tooltip is generated rather than written and cannot drift from the query. Grafana has no notion of a metric definition at all—panels carry a description field, and whether it holds a real definition or nothing is a matter of author discipline. Honeycomb sidesteps some of it by keeping raw events: the derivation is visible in the query rather than hidden behind a metric name someone chose months ago. Sentry has an advantage most products lack, in that its core objects are concrete enough to need little definition; an error either happened or it didn’t, and the ambiguity lives in grouping rules rather than in counting.
Related
Data source badge answers where the number came from, which is the other half of the same trust question. Ratio and rate is the pattern where a hidden denominator does the most damage. Metric targets is where a definition becomes load-bearing, because a target on an ambiguous metric is an argument waiting to happen. KPI tile is the container that most needs this and most often lacks it. Hover detail is the interaction, and the entry that warns against making a tooltip carry a paragraph.