Charts That Answer the Question You Asked

Chart selection is usually treated as a matter of taste or variety. It is neither. The right chart follows mechanically from what you want the audience to compare, and lesson one already established what that is: the message in your sentence title.

The mapping, which covers most business situations.

Change over time: a line chart. Time on the horizontal axis, always, because readers expect it there and violating that costs comprehension for no gain.

Comparison between categories: a bar chart, sorted by value rather than alphabetically unless the categories have a natural order. Sorting is the single cheapest improvement most charts can receive.

Composition, where parts make a whole: a stacked bar, or several bars showing each part separately, which is usually clearer.

Relationship between two variables: a scatter plot.

Distribution: a histogram, or a box plot when comparing distributions across groups.

Deviation from a reference: a bar chart with bars going both directions from zero.

What is missing from that list is instructive. Pie charts appear for composition and are worse than bars for almost every purpose, because people compare angles poorly. And there is no entry for making it more interesting, because that is not a comparison anyone needs to make.

The practical test. Write the sentence title first, then ask what the reader must compare to believe it. The chart type falls out of the answer, and if no comparison is required, you do not need a chart.

The reason some charts work better than others is not convention. It is that human perception judges some visual properties far more accurately than others, and this has been studied.

William Cleveland and Robert McGill published experiments in the Journal of the American Statistical Association in 1984 testing how accurately people judge quantities encoded in different ways. Their ranking, which has held up and been extended since, runs roughly as follows.

Most accurate: position along a common scale. Two bars on the same axis, two points on the same line.

Then position along non-aligned scales, as in separate small charts side by side.

Then length, angle, and area, in declining accuracy.

And least accurate: colour saturation and shading, where people can reliably distinguish only a few levels.

What follows practically.

Bars and dots on a shared axis are the workhorses because they use the most accurate encoding available.

Pie charts use angle and area, which is why comparing slices is hard, and why they only work for the case where one slice is obviously dominant.

Bubble charts, where quantity is encoded as circle area, are systematically misread, and worse, many tools size bubbles by radius rather than area, which exaggerates differences dramatically.

And heatmaps encode with colour, the weakest channel, so they are good for spotting patterns and bad for reading values.

The general principle. If a number matters, encode it as position. If it is context, colour is acceptable. And if you find yourself needing a legend to decode saturation levels, the chart is asking the reader to do something people do badly.

The decision path, starting from the sentence title rather than from the data.

Begin with the message you wrote. What must the reader compare to accept it?

If the comparison is across time, a line chart, with time horizontal.

If it is between categories, a bar chart, sorted by value. If the categories have a natural order such as age bands, keep that order instead.

If it is between parts of a whole, prefer separate bars over a stacked bar, and prefer either over a pie unless one part obviously dominates.

If it is between two measures, a scatter plot.

If it is about spread rather than a central value, a histogram or box plot.

And if the answer is that no comparison is needed, because you are stating a single fact, then the right exhibit is a sentence with the number in it. A chart of one number is decoration, and a large number in text communicates it better.

That last branch is used far too rarely. A great deal of what appears as charts in business presentations would land harder as a single line of text, and reaching for a chart by default is one of the more common ways presentations get longer without getting clearer.

Whether a bar chart's axis must start at zero is the most-argued question in this field, and it has a clear answer that depends on the encoding.

Bar charts must start at zero. A bar encodes quantity as length, so the reader is comparing lengths. Truncating the axis makes a bar representing 98 look half the size of one representing 100, which is not a matter of taste. It is a false statement expressed visually, and it is the single most common way charts mislead.

Line charts need not. A line encodes change as slope and position, and the reader is judging the shape of movement rather than comparing lengths from a baseline. Forcing a line chart of a stock index to start at zero flattens the entire signal into a horizontal line, which destroys the information.

So the rule follows from the encoding: length-based charts need the zero, position-based charts do not.

What still applies to line charts is honesty about scale. A y-axis chosen to make a two percent fluctuation fill the frame will read as a crisis, and while it is not technically false, it is misleading in effect. The defence is to label clearly and to choose a range that reflects what variation is meaningful in this context.

Two related manipulations worth recognising, in your own work as much as in others'.

Dual axes, where two series with different scales share a chart, can be made to appear correlated by choosing the scales appropriately. They are almost always better as two charts.

And selective time ranges. Starting a trend at a convenient point is the most effective way to mislead with entirely accurate data, and it is worth asking yourself why you chose that start date.

Most business charts carry a substantial amount of material that competes with what they are trying to say, and removing it is the fastest available improvement.

Edward Tufte's work on graphical display, particularly The Visual Display of Quantitative Information published in 1983, framed this as the ratio of ink used to display data against ink used for everything else. The principle is that decoration competes with information for the reader's attention.

What usually goes.

Gridlines, or at least most of them. Heavy grids draw the eye to the background.

Borders and background fills.

Three-dimensional effects, which add no information and actively distort the reader's judgement of length and area.

Redundant legends. If there are two lines, label them directly on the lines. A legend forces the reader to look back and forth, which is a real cost repeated for every value they read.

Excessive decimal places. Revenue of 4,238,917 is harder to read and no more useful than 4.2 million, and precision beyond what the data supports is misleading about certainty.

And colour used decoratively. If every category has its own colour but the colours mean nothing, they consume attention without carrying information.

What colour should be used for instead: directing attention. Grey everything, then colour only the series your sentence title is about. This single technique does more for a business chart than any other change, because it makes the point visually rather than requiring the reader to find it.

The test. Does every element on this chart help the reader believe my title? If not, remove it.

A category of error that no chart design fixes, because the problem occurs before the chart: aggregating away the finding.

An average describes a distribution badly whenever the distribution is not roughly symmetric, and business distributions rarely are. Revenue per customer, time to resolution, session length and order value are typically skewed, with a long tail that pulls the mean away from anything typical.

The practical consequences.

Average revenue per customer of two hundred can describe a population where most customers spend forty and a handful spend thousands. The average describes nobody, and a decision made on it will be wrong for almost everyone.

An overall figure that has not moved can conceal two segments moving in opposite directions, which is the case where the aggregate actively misleads about whether anything is happening.

And a general improvement can coexist with a decline in every subgroup, or the reverse, when the mix between groups changes. This is Simpson's paradox, and it is not a curiosity: it appears routinely in real business data whenever the composition of a population shifts.

What to do about it, practically.

Look at the distribution before presenting a summary of it, every time. A histogram takes seconds and frequently changes the finding.

Present the median alongside or instead of the mean where the distribution is skewed, and say which you used.

Break out the segments that matter before concluding, particularly when an aggregate shows no change.

And when you present an average, show the spread. A bar showing a mean with no indication of variation is technically accurate and routinely produces a wrong decision.

Dashboards get built with presentation instincts and then fail, because they are answering a different question.

A presentation makes an argument to a specific audience about a specific decision, once. A dashboard supports monitoring by many people over time, and nobody is there to explain it.

What follows from that difference.

A dashboard needs to answer is anything wrong, at a glance, and most do not. A grid of twenty charts requires the viewer to assess each one, which is work, so they stop looking. The fix is to make the normal state visually boring and the abnormal state visually loud, so attention is drawn rather than solicited.

Every metric needs its comparison built in, since there is no presenter to supply context. A number without a target, a trend or a prior period is uninterpretable, and a dashboard full of bare numbers is a dashboard nobody can act on.

Definitions must be visible. When many people read the same metric over months, they will interpret it differently unless the definition is attached, and disputes about what a number means consume more time than producing it did.

And the failure mode mirrors alert fatigue from the supply chain cursus. A dashboard showing everything gets ignored exactly like an alert system that fires constantly, and the ignoring is worse than absence because the organisation believes it is monitoring.

The design question that matters. What will someone do differently based on this? A dashboard whose metrics have no associated action is a report that updates itself, and it should be a report.

What to run through before an exhibit goes in front of anyone.

Does it have a sentence title stating the message rather than describing the axes?

Does the chart type follow from the comparison that message requires? Line for time, sorted bars for categories, scatter for relationships, and a sentence rather than a chart if no comparison is needed.

If it uses length, does the axis start at zero? If it uses position, is the range honest about what variation matters?

Is the reader being asked to judge anything by area, angle or colour saturation? Those are the encodings people read worst, and anything important should be position.

Is colour doing work? Grey the context, colour only what the title is about.

Have I removed the gridlines, borders, three-dimensional effects, redundant legends and spurious decimal places?

Did I look at the distribution before summarising it, check whether segments move differently from the aggregate, and show spread alongside any average?

And would a competitor or a sceptic looking at this find a choice that flattered me? The truncated axis, the convenient start date, the dual axis, the mean rather than the median. Each is a choice, and an audience that later spots one will discount everything else.

The unifying idea across both lessons. The chart is not the output. The decision is. Everything in the exhibit exists to help one person believe one claim well enough to act, and anything not doing that is in the way.