A useful chart is a decision aid. It helps someone answer a question: which region is underperforming, whether an intervention worked, where the budget went, which customers are changing behaviour, or what risk sits outside the expected range.

The starting point is not "what chart should I use?" or "what would look impressive?" The starting point is why the information needs to exist.

A practical design process

Five stages make the work explicit: define the purpose, establish the evidence, choose an effective visual encoding, build a clear hierarchy, and test the result in use.

They are not a production checklist to be completed mechanically. Each stage constrains the next. A change to the decision can change the evidence required; a change to the evidence can change the chart; testing may reveal that the original question was too broad.

1. Define the purpose

Start with the reason the chart exists. "We need a dashboard" is not a purpose. Neither is "the board wants a pie chart". Those are requests for an output.

A useful purpose names the reader, the question and the decision:

  • Who will use it? An executive monitoring risk, a manager allocating resources and an analyst investigating causes need different levels of detail.
  • What question must they answer? Which region is underperforming? Is the change outside the normal range? Where is conversion being lost?
  • What decision, judgement or action follows? Reallocate budget, investigate an exception, change a target, or simply maintain awareness.
  • How often will the question be asked? A one-off presentation, a monthly report and an operational display are different products.

If the purpose is unclear, establish it before designing anything. Otherwise the work begins with a visual preference and searches for a reason to justify it.

2. Establish the evidence

The chart is only as sound as the evidence it represents. Before choosing a form, define what must be measured and what context makes the measure interpretable.

Ask:

  • What exactly does the metric mean?
  • What is the denominator, population or unit of analysis?
  • What time period and level of detail are appropriate?
  • What comparison makes the value meaningful: previous period, target, forecast, peer group or expected range?
  • Does aggregation hide variation, outliers or unequal sample sizes?
  • What uncertainty or data-quality limitation needs to remain visible?

A total without a denominator can reward scale rather than performance. An average without a distribution can conceal two very different populations. A forecast without an interval presents uncertainty as false precision. Better styling cannot repair these problems.

This stage also determines whether a chart is the right form at all. Use a table when readers need exact lookup across several fields. Use a chart when the task is to see shape, pattern, comparison, exception or change. Often the strongest answer combines a chart for the pattern with a small table or direct labels for exact values.

3. Design for visual perception

Charts encode values into visual properties: position, length, angle, area, colour, shape and movement. People do not read those properties with equal accuracy.

Research into graphical perception consistently finds that position on a shared scale supports more accurate comparison than length, while angle and area are less precise. Colour is powerful for grouping and emphasis, but poor for reading exact magnitude. Chart recommendations are grounded in those differences; they are not arbitrary conventions.

Why the usual first choices work

Use bars or dots for comparison. Values share a common baseline or scale, so the eye can compare position or length directly. A horizontal bar also gives long category labels room to remain readable. A pie chart can show that a few parts form a whole, but it asks the reader to compare angles and areas. When the question is "which is larger, and by how much?", a bar or dot plot is the stronger instrument.

Use lines for continuous change. The connecting path turns observations into a trajectory. Direction, rate of change and turning points become visible as shape. Use columns when periods are discrete or when absolute period totals matter more than continuity.

Use distributions instead of averages when variation matters. Histograms, density plots, box plots and point-based distributions reveal spread, skew, clusters and outliers that a bar of means removes. A summary statistic is useful only when the variation it summarises is not the story.

Use bullet graphs for performance against a target. A bullet graph places actual performance, a target and useful reference ranges on one compact linear scale. A speedometer-style gauge consumes far more space, usually displays less context and asks the reader to judge an angle along an arc. Familiarity does not make that encoding efficient. This is a perceptual and informational problem, not an aesthetic objection.

Use small multiples when several series need equal attention. Giving each series the same scale in its own panel preserves comparison without forcing readers to trace a dense bundle of lines through a legend. Use a highlighted multi-series chart when one series is the subject and the others are context.

Use direct labels when they remove decoding work. A label beside a line, bar or notable point is usually easier to follow than a legend that requires repeated lookup. Legends remain useful when direct labels would collide or when many marks share a stable categorical scheme.

The aim is not to apply a universal ranking without judgement. It is to use the most accurate visual encoding that fits the question, the evidence and the available space.

4. Build hierarchy and context

Once the chart form is right, design the reading order. The most important evidence should be easiest to find; supporting context should remain available without competing for attention.

  • Write a clear title that identifies the subject. Use the subtitle for units, dates, population and comparison context.
  • Label important values, series, targets and exceptions directly where practical.
  • Keep scales honest and comparable. Bars showing absolute magnitude normally begin at zero; focused line-chart scales should be clearly labelled.
  • Use reference lines and ranges when the reader needs to judge performance, normality or risk.
  • Remove decoration that does not carry information, but retain scaffolding that helps the reader judge the data.
  • Use annotation to explain consequential events or changes, not to narrate every point.
  • Preserve a visible source and explain material caveats.

The data-ink principle is useful here, but it is often misunderstood. The goal is not to make every chart sparse. The goal is to remove ink that competes with the evidence while retaining the labels, scales, comparisons and context needed to understand it.

Colour is part of this hierarchy, but it is not a finishing treatment. Inside a chart, colour should identify, order, emphasise or signal. The palette families and accessibility considerations are covered in Part III — Colour and Data Palettes.

5. Test readability in use

A chart is not finished when it renders correctly. Test whether people can use it to answer the intended question.

Ask representative readers to look at the chart in its real context and explain:

  • What do you notice first?
  • What comparison or pattern do you see?
  • What does the chart imply?
  • What would you do next?
  • What, if anything, is difficult to interpret?

For a monitoring chart, the key status and exception should usually be apparent within seconds. An exploratory analytical chart may reasonably take longer, but its controls, encodings and reading path must still be learnable. "It is a complex subject" is not permission for avoidable complexity.

Test at the sizes and conditions in which the chart will be used: a laptop, a mobile screen, a projected presentation, a printed board pack or a dense operational display. Check labels, interaction states, hover content, keyboard access, colour, contrast and what remains visible without interaction.

Readability is task-based. A design succeeds when readers can discern the evidence required to make an informed judgement with appropriate speed and accuracy.

Preference is not evidence

Preference, branding and communication are often confused.

"I like gauges" is a preference. "Use our corporate colours" is a branding request. "Can the reader understand performance against target and decide whether to intervene?" is a design question.

Preferences and brand systems can influence the presentation, but they do not override how people perceive quantitative information. When a decorative treatment makes the evidence harder to compare, the evidence should win.

Reading list