2026-09-15
Nural Choudhury
Information design is the craft of presenting information so it can be read and reasoned about: chart and encoding choice, hierarchy, annotation, and the honest representation of quantity.
Structuring content so a reader can find what they need is a different job, and information architecture covers it.
the ability to choose an encoding that shows what is true, and to catch a chart that is technically accurate and still misleading, before it ships.
any chart, table, dashboard, or diagram put in front of a reader who has to compare a quantity, follow a change, or make a decision based on what they see.
state the question before you choose the chart type. A chart picked out of habit, from a template, or because it’s the tool’s default answers no question.
“In The Visual Display of Quantitative Information (1983), Edward Tufte argued that every mark on a graphic should carry data, and called the marks that do not carry data. The ratio of data-carrying ink to total ink is the data-ink ratio. The argument still governs how a working information designer edits a chart: remove what does not carry data, then check that what is left is still legible.
The vocabulary for what a mark can encode is older. Jacques Bertin’s Sémiologie graphique (1967) set out the visual variables, the properties a mark can vary to represent a value: position, size, shape, value, colour and orientation. Later research ranked these variables by how accurately a reader can decode them, and that ranking is why the table below starts with position and ends with colour.
The chart itself is newer than the habit of counting. William Playfair introduced the line and the bar chart in The Commercial and Political Atlas (1786), and the pie chart in the Statistical Breviary (1801), because a table of numbers hid a trend his readers needed to see at a glance.
Every chart makes three decisions: what mark encodes the value, what the reader sees first, and what the reader is told to notice. Get the first wrong, and the chart lies by construction. Get the other two wrong,g and it lies by omission.
| Encoding | What it conveys well | Where it misleads |
|---|---|---|
| Position along a common scale | The most accurate encoding a reader can decode; a difference in position reads as a difference in value | Rarely on its own; a moved axis is the usual cause of a misleading position chart |
| Length | Magnitude from a shared baseline, as in a bar chart | Once the baseline is not zero, because the reader compares bar lengths, not the values behind them |
| Angle | Ordering a small number of shares of a whole | Once more than a handful of segments compete, because a reader judges a wide angle less accurately than a narrow one |
| Area | Rank order among values with a wide range | Badly, when a design scales the radius rather than the area, because the visible size then grows with the square of the value |
| Colour (hue) | A category, not a magnitude | Whenever hue is asked to imply order or size, which a reader cannot rank by eye |
Hierarchy is the order in which a reader’s eye moves through a graphic, and you set it with scale, contrast, colour and position, not with a numbered list beside the chart. Put the reading you want a viewer to reach first at the largest scale, the highest contrast, or the top-left of a Western reading pattern, and put supporting detail in a lower layer behind it. General composition, balance and white space across a full page is not this page’s job; see layout principles, since the hierarchy taught here sits inside one graphic.
An annotation names what the reader should notice; a caption that repeats the axis label does not. Write the annotation as the sentence you would say out loud if you were standing next to the reader, pointing at the feature that matters. Put that sentence on the chart itself, not in a paragraph beside it.
Graphical integrity is not a matter of taste. A truncated axis turns an ordinary change into what looks like a trend, because the reader compares bar heights rather than the numbers underneath them. Scale a bubble or an icon by radius rather than by area, and the same fault appears twice over, because a doubled value then produces a shape with four times the visible area.
A second axis on the same chart lets two unrelated series appear to move together purely because their scales happen to line up. The correlation is an artefact of the axis choice, not of the data.
The harder failures are the ones where every number on the chart is correct. A time series that begins the week after a spike is accurate and still hides the spike. A monthly aggregate that smooths out one bad week is accurate and still misrepresents what happened that week.
A chart that shows a rate without showing the base it is a rate of is accurate and still lets a small absolute number look dramatic. None of these breaks a rule you can check with a ruler. Each one requires you to ask what the chart leaves out, not only what it shows.
How to build the palette itself, the wheel, the scheme, the harmony, is not this page’s job; see colour theory basics. Whether that palette reads at a legal contrast ratio, and whether a reader with a colour vision deficiency can decode it without relying on hue, is not this page’s job either; see accessibility standards.
State the question the chart must answer before you open a charting tool. “How has this changed” and “how does this compare” are different questions with different correct encodings. A chart built before you settle the question answers neither well.
Work down the accuracy order in the table above: position before length, length before angle, angle before area, and area before colour. Choose the least accurate encoding only when the more accurate ones will not fit the data or the space.
You must start a length or area encoding at zero. Where a non-zero baseline genuinely serves the reader, you must label it on the axis itself, not in a footnote.
You must hold the scale constant across every chart in a set the reader is meant to compare directly. A dashboard where one tile runs from zero to a hundred and the next runs across a narrow middle band invites a false comparison even when both are individually accurate.
Write the annotation before you ship the chart, not after a reviewer asks what it means. If the chart moves or animates, the timing that keeps a reader oriented while it changes is not this page’s job; see animation principles. Get the static version honest first.
Where the data itself needs interrogation, cleaning, or a choice of statistical method before it reaches a chart, that is not this page’s job; see data analysis for design. Bring a clean, well-understood dataset to the decisions on this page.
A retail client’s quarterly report showed revenue across several regions as a bar chart with an axis that started well above zero, because the regions clustered close together and a zero baseline made the bars look nearly identical. The chart was accurate. It also made an ordinary spread look like a stark gap between the best and worst region, which was not the story the underlying numbers told.
I rebuilt it at zero, which flattened the bars and let the real story- that the regions were close- read from the chart itself rather than from an explanation attached to it. I added one annotation naming the two regions the report needed the reader to compare, instead of a caption repeating the axis units. The redesigned chart was less dramatic and closer to the truth, which is the trade the client needed to see spelt out before they accepted it.
I have watched a chart type get picked before anyone in the room stated the question, usually because a template or a default in the charting tool supplied one automatically. The result answers a question nobody asked, and the meeting spends longer arguing about the chart than about the data.
I have seen a pie chart carrying eleven segments, most of them slivers a reader cannot compare by angle at any size. Past four or five segments, a pie chart asks the reader to do arithmetic in their head. A sorted bar chart does the same job and asks for nothing.
I have seen a truncated axis turn a rounding error into what a reader read as a trend, in a chart nobody intended to mislead. The axis choice was habitual, carried over from the last report, not checked against the actual spread of that quarter’s numbers.
I have seen an annotation that repeats the axis label rather than naming what changed, on the theory that any text on a chart counts as context. It does not. An annotation earns its place only if it tells the reader something the axis and the shape of the data do not already say.
Why does area mislead more often than length? A reader compares two lengths reasonably accurately, but area grows with the square of a linear change. A design that scales a shape’s radius rather than its area exaggerates every difference twice over, whether anyone intends it or not.
Should an axis always start at zero? For a length or area encoding, yes. A position encoding, such as a line chart tracking a rate over time, can start elsewhere because the reader reads the slope, not the length of a bar against a baseline. Label the choice either way.
When is a pie chart the right choice? For two or three segments, where the reader only needs to see that one share is clearly bigger than another. For four or five segments, a sorted bar chart conveys the comparison more accurately.
What is chartjunk? Any mark on a graphic that does not carry data: a three-dimensional bevel, a decorative background, a gridline denser than the data needs. Removing it is usually the fastest improvement you can make to an existing chart.
| Item | Where it stands |
|---|---|
| Data-ink ratio | Introduced by Edward Tufte in The Visual Display of Quantitative Information (1983); still the working argument against chartjunk |
| Visual variables | Jacques Bertin’s Sémiologie graphique (1967) set out position, size, shape, value, colour and orientation as the properties a mark can vary to encode a value |
| Origin of the statistical chart | William Playfair introduced the line and bar chart in The Commercial and Political Atlas (1786), and the pie chart in the Statistical Breviary (1801) |
Information architecture structures content so people can find it: choosing an organisation scheme, writing labels, building navigation and search, and testing whether the structure holds under a real task.
Read it
Colour theory is the study of how colours relate to each other on the wheel, across value, saturation, and temperature, and in combination so that you can reason through a decision rather than feel it.
Read it
Digital accessibility is the practice of building products people can use, whatever their permanent, temporary, or situational disability, measured against a published standard rather than opinion.
Read it
Animation principles are the small set of rules, developed for hand-drawn character animation, that make movement read as caused by weight and intent rather than as an arbitrary jump between two states.
Read it
Data analysis for design is the practice of testing decisions against measured evidence, usage data, and controlled experiments, rather than defending them by opinion alone.
Read it
Layout is the arrangement of elements inside a frame: what a viewer sees first, second, and third, and what the space between things does as they look.
Read it