Build brilliant Power BI reports with a disciplined schematicLesson 5 of 8
Picking the right visual
Use a selection matrix to match native visuals to questions and qualified fields.
What you'll do
Choose native visuals for a synthetic report using the question, field grain and available evidence. Build a selection matrix and reject one misleading alternative. You need existing measures and related dimensions; this lesson does not create missing measures or establish their correctness.
Match the visual to the question
Begin with the comparison a reader must make. Time, rank, composition and record lookup are different tasks. A visual's popularity is not a reason to use it. Prefer a small set of native types whose roles and behaviour can be checked against the installed catalogue.
| Question | Default visual | Evidence or restriction |
|---|---|---|
| Current result versus baseline | kpi, then cardVisual, then card | Qualified endpoint/baseline semantics |
| Change over time | lineChart | Real chronology and explicit sort |
| Categories leading or lagging | clusteredBarChart | Ranked measure and stable ties |
| Volume and rate together | lineClusteredColumnComboChart | Paired question, named units and axes |
| Parts of a whole | barChart or hundredPercentStackedBarChart | Exclusive additive parts or qualified shares |
| Reconciled change | waterfallChart | Start, end and additive bridge inputs |
| Association | scatterChart | Profile evidence and comparable entity grain |
| Distribution | clusteredColumnChart | Governed numeric bins |
| Exact values or hierarchy | pivotTable | Correct totals and useful grouping |
| Actionable records | tableEx | Predicate and severity/exposure sort |
These names describe intended native PBIR types. Native columnChart means stacked columns; native barChart means stacked bars. A logical alias such as stackedColumnChart is not itself a native serialization name. Check supported role mappings rather than guessing field roles.
Make ordering part of the specification
Time ascends chronologically. Category rankings use impact or value, normally descending. Aging bins and process stages follow business order. A month name sorted alphabetically is a semantic error in the presentation even when every underlying value is correct.
Bar and column axes start at zero. A line can use a clearly labelled narrower range. Comparable small multiples share scales. Aim for a readable number of ticks, horizontal labels and explicit units. Never hide an ambiguous second axis behind a plausible-looking combination chart.
Synthetic ranking contract:
Measure: qualified sales amount
Category: business territory label
Order: amount descending, stable territory key for ties
Coverage: Top 10; Other only if independently reconciledThe specification names business labels rather than surrogate keys. It also identifies how the omitted population will be explained. A Top-N chart must not imply that the visible members are the complete business population.
Be careful with composition
Other must represent the excluded set. For exclusive additive amounts, reconcile the visible sum plus Other to the total. For overlapping distinct counts or rates, recompute over the excluded population or omit Other and disclose coverage. Subtracting displayed percentages is not an independent definition.
A waterfall needs additive bridge contributions that reconcile its start and end. A set of unrelated totals or ratios cannot become a bridge because a waterfall looks persuasive. Similarly, a scatter plot needs evidence of comparable entity grain and valid size/cardinality. An observed association does not prove causation.
Review exact detail and totals
Use Matrix when hierarchy or row grouping helps. Apply data bars to one qualified numeric measure rather than every column. Use a flat table for an action queue. Re-evaluate ratios and distinct counts at the total's scope; do not average visible rates or add overlapping populations to mimic a total.
Finish your matrix with unavailable-state behaviour. If there is no valid time field, choose a rank instead of inventing a trend. If the model cannot answer the page's question, suppress the page and update navigation. Your expected output is a selection rationale, sorting contract and evidence list. Preview v2 authoring remains subject to the installed catalogue; visual selection alone does not qualify queries, interactions or rendered labels.
Check your understanding
- What independent condition makes a waterfall appropriate?
- Why is Other unsafe when calculated from rounded visible percentages?
- Which ordering and axis choices belong in a visual specification?