The chart you reach for is an argument, not a default
Every chart type encodes a claim about your data — and reaching for the wrong one doesn't just look bad, it actively misleads the people making decisions.
In 2012, a trading desk at JPMorgan lost roughly 6.2 billion dollars. The internal risk models that were supposed to catch the position used spreadsheets — spreadsheets with hard-coded values instead of live feeds, and summaries that aggregated the exposure into a single blended number. Nobody drew a scatter plot of position size versus volatility. Nobody charted the trend in daily mark-to-market swings. The warning signals were in the data. The data was never put into a shape that made them visible.
Charts are not decoration. They are arguments. And a chart that makes the wrong argument — or no argument at all — is the same as silence, except more expensive.
The claim hidden inside the choice
Every chart type encodes a specific assertion about the relationship between your variables. This is the part nobody teaches explicitly, and it is the source of most bad visualisation in industry.
A line chart asserts: this value changed meaningfully over time, and the trajectory matters. The connected line implies continuity — that the value at Tuesday is causally related to the value at Wednesday. Use a line when the x-axis is time and the in-between points are real, not gaps.
A bar chart asserts: these discrete categories have magnitudes worth comparing side by side. The visual unit is height (or length in a horizontal bar), and humans read height comparisons with high accuracy. The categories are not connected — there is no implied story between bar three and bar four.
A scatter plot asserts: these two continuous variables have a relationship — look at its shape. The scatter is the only chart that shows correlation (a statistical measure of how much two variables move together) without summarising it away. It shows the outliers, the clusters, the funnel shapes that aggregated statistics always erase.
A single big number — just a KPI (key performance indicator: a single headline metric) on a card — asserts: one truth matters right now, and everything else is context. It is the most opinionated of all charts. It refuses to show distribution, trend, or comparison. That refusal is a choice.
When you reach for a chart type out of habit, you are not picking a neutral container. You are endorsing one of those claims. If your data does not support the claim, your chart lies — quietly, visually, in a way that slides past the critical faculties of anyone who was not already suspicious.
Data-ink and the cost of decoration
Edward Tufte, the statistician and designer who wrote The Visual Display of Quantitative Information in 1983, introduced two ideas that have outlasted most of their era: data-ink ratio and chartjunk.
Data-ink is every pixel on the page that represents an actual data point. Chartjunk is every pixel that does not: the thick grid lines, the gradient fills, the 3-D bevels on bar charts, the drop shadows, the legend boxes that duplicate labels already present in the chart itself.
The data-ink ratio is the fraction of a chart’s total ink that is doing informational work. Tufte’s argument — and decades of cognition research have supported it — is that maximising this ratio reduces cognitive load and increases the speed and accuracy with which readers extract the intended message. A chart drowning in gridlines forces the eye to work harder to find the bars. A chart with a subtle, thin baseline and no gridlines at all lets the bars speak.
The practical implication for anyone building dashboards or slide decks: every element you add to a chart that carries no information costs the reader something. A 3-D pie chart does not add depth — it distorts the areas that are already the primary visual encoding. An axis that starts at 60 instead of zero does not add detail — it makes a 5 percent difference look like a 400 percent difference. These are not neutral stylistic choices. They are quiet arguments made on behalf of whoever drew the chart.
The question to ask about every element of every chart is simple: if I removed this, would the reader lose information? If the answer is no, remove it.
Why pie charts almost always fail
The pie chart is the most defended bad idea in data visualisation. Defenders invoke familiarity. They point out that audiences understand what a pie chart means. They are right about that. They are wrong about nearly everything else.
Humans read area and angle comparatively poorly. The perceptual research — Eells (1926), Cleveland and McGill (1984), and replicated many times since — consistently shows that human judgment of relative lengths is accurate to within about 5 to 10 percent, while judgment of relative angles is accurate to only about 20 to 30 percent under the same conditions. A bar chart of the same categories as a pie chart will be read correctly more often, more quickly, and by a wider audience.
The pie chart compounds this by making comparison across charts nearly impossible. If you want to show how a market split across segments changed from 2023 to 2025, you need two pies side by side. Two pies are nearly unreadable as a comparison because the starting angles of the slices are different. Two grouped bar charts are effortless.
The one scenario where a pie chart is defensible: when you have two or three categories and your only message is the proportion of the whole. “We get 80 percent of revenue from five customers and 20 percent from everyone else.” A single donut chart (a pie with a hole in the centre, which forces comparison by arc length rather than area) is acceptable. Anything with more than five slices has lost the argument already, because the audience will spend their cognitive budget trying to read thin sliver differences that are better expressed as a table.
One chart, one message
The failure mode that senior analysts fall into — and that data teams at fast-growing companies reproduce at scale — is the chart that tries to answer three questions at once.
Put revenue, user count, and gross margin on one y-axis (a scale used to read values off a chart), differentiated by colour. Add a secondary y-axis on the right for a fourth metric. Add a trend line. A reader looking at this is not enlightened; they are doing document parsing. They are scanning for the one number they actually need and ignoring the rest.
A chart that requires a paragraph of explanation in the slide notes has failed. That is not a standard to aspire to; it is a test to fail quickly and revise.
The discipline of one chart, one message does not mean the chart is simple. A well-designed scatter plot with 10,000 points and a visible cluster structure carries enormous information density. But it carries one message: look at the shape of this relationship. Every data point earns its place. No point is decoration.
When you catch yourself thinking “I’ll add a second metric so the context is there,” ask instead whether the context deserves its own chart. Two charts on one page, each making one clear argument, are almost always more readable than one chart making two arguments badly.
The decision you are actually making
When a product manager sends a weekly metrics email and picks a line chart for seven-day rolling retention, they are telling their readers: the trend matters more than this week’s number in isolation. When a finance analyst puts revenue by quarter in a bar chart rather than a line, they are saying: these quarters are discrete achievements to compare, not a continuous trajectory to follow. When an ops team pins a single 99.4% uptime number on a status page, they are saying: everything else about our infrastructure is irrelevant to you right now; trust this one truth.
These are editorial decisions. They are closer to writing a headline than to rendering a spreadsheet. The chart type is the framing, and framing changes what people believe.
This is why a chart that misleads is almost never the result of someone deliberately lying with data. It is the result of someone choosing a chart type out of habit, adding the decoration their company template supplies by default, and never asking what claim they are making. The lie is not intentional. It is just unexamined.
Matching chart to claim
The mental model that fixes most chart choices is to start with the sentence you want the reader to finish when they look away from the chart. Not “this shows revenue over time.” The actual message: “Revenue has grown steadily but Q4 is consistently our strongest quarter, and last year it underperformed for the first time.” That message calls for a bar chart grouped by quarter with year-over-year comparison — not a line, because the quarterly rhythm matters more than the continuous trend.
The diagram below maps the most common message types to the chart that best encodes them, and flags the chart types that are commonly reached for but wrong.
The second discipline the diagram points at: notice that the “wrong” choices are not random. They are each systematically wrong in the same direction. A bar chart used to show a time trend suppresses the trajectory. A pie chart used to compare categories obscures the magnitudes. A line chart used to show a distribution implies that the values are ordered and connected, which they are not. The mistakes are predictable. That means they are fixable.
Chartjunk in the wild
Open any major consulting firm’s published slide deck and run the Tufte test. You will find: 3-D bars where the depth carries no data. Gradient fills on pie slices that make the near slices look lighter, biasing area perception. Tick marks on both sides of an axis. Legend boxes in the middle of the chart area, breaking the data field. A company watermark on every chart that is never relevant to the data.
None of these features are malicious. They are the output of PowerPoint and Tableau defaults, templates built for branding rather than communication, and a cultural norm that “more polish” means “more elements.” The result is charts that are aesthetically elaborate and cognitively expensive — the opposite of what a chart is for.
The corporate dashboard is a particularly rich environment for chartjunk. The typical Tableau dashboard (Tableau is a business intelligence tool for interactive visual analytics) has at least four chart types on one screen, a colour palette that repeats the same three corporate blues across unrelated metrics, and a date filter that applies to some charts but not others without any indication of which ones. A viewer trying to extract insight from this has to hold a mental model of the filter state, the scale of each axis, the unit of each metric, and the relationship between the charts. This is not data communication. It is data archaeology.
The discipline that cures this is not design training. It is the habit of asking, before publishing any chart: what is the one sentence I want someone to say after they see this? If you cannot write that sentence, the chart is not ready.
The data-ink calculation in practice
A clean bar chart has: bars, a single y-axis with minimal labels, a single x-axis with category labels, and optionally a title. That is it. Every additional element needs to earn its presence.
Do the horizontal grid lines help? Only if the values are hard to read from the bar heights alone — which is often true when bars are narrow or the differences are small. Use light, thin gridlines and only enough of them to anchor the scale. Do not put them on every round number.
Does the legend need a box? No. A legend box is a container for the legend labels. The labels are the information. The box is decoration.
Does the y-axis need to start at zero? For a bar chart, almost always yes. Bar height encodes magnitude by area. Cutting the axis at 60 makes a bar representing 65 look five times taller than one representing 61, when the actual difference is just under 7 percent. For a line chart showing change over time, starting at zero is sometimes wrong — if the range of variation is small relative to the absolute value, a zero baseline compresses the signal into a nearly flat line and destroys the message.
The axis baseline rule is another example of the same principle: every visual choice is an argument, and the argument needs to match the claim.
A practical test
Before publishing any chart, apply three tests in order.
First: cover the title. Can a smart peer who works in your domain still tell what the chart is arguing? If not, the chart is not making its case through the visual encoding. The title is covering for a failure of design.
Second: describe the chart in one sentence starting with “this shows.” If the sentence requires a semicolon or a subordinate clause, you are probably making more than one claim. Split it.
Third: list every non-data element — gridlines, legend, tick marks, background, border, drop shadow. For each one, ask whether removing it would cost the reader information. Remove the ones that cost nothing.
These tests take about ninety seconds. They consistently produce cleaner, faster, more persuasive charts than another pass through the colour palette.
Charts as institutional memory
There is one more argument for taking chart type seriously that goes beyond the immediate communication: charts are how decisions get revisited.
A post-mortem (a retrospective analysis of what went wrong) on a product decision made six months ago will often open with a chart from the original analysis. If that chart made the wrong argument — if it showed a trend where there was noise, or hid variance behind a mean, or cut the axis to dramatise a small difference — then the post-mortem will spend the first twenty minutes arguing about the chart instead of the decision. The chart did not just mislead the original audience. It corrupted the institutional record.
This is why firms like McKinsey and Goldman Sachs spend an unusual amount of time on chart standards. Not because they are aesthetes. Because charts are the primary medium through which evidence survives a hand-off. When the analyst who built the model leaves, the chart is what the new team reads. The encoding choices made in a Tuesday afternoon Tableau session outlast the person who made them.
Choosing a chart type is not a finishing step. It is a claim about what matters in your data, how it should be compared, and how certain you are. Make the claim deliberately, strip out everything that is not the claim, and write the one sentence that should be in the reader’s head when they look away.
If you can do that consistently, you are doing something most analysts never manage: communicating instead of decorating.