How to choose the right chart for your data

Start with the question, not the data

Write the question in a sentence before choosing anything. The verb tells you the chart type:

• "Which region sold most?" — a comparison.
• "Is churn going up?" — a trend.
• "What share of revenue is subscriptions?" — a composition.
• "How are order values spread?" — a distribution.
• "Does discount affect margin?" — a relationship.

If the sentence contains two verbs ("how has each region’s share changed?") you need two charts or a carefully chosen one. If you cannot write the sentence, you are not ready to pick a chart. The rest of this guide covers each verb in turn; the glossary entry on data visualisation covers the vocabulary.

Comparing categories: bar charts

A bar chart is the right answer for almost every comparison, and it is under-used because it looks plain. Rules of thumb:

1. Horizontal bars when the labels are words (product names, regions, reps). Vertical bars when labels are short or ordered (months, sizes).
2. Sort by value, largest first, unless the categories have a natural order. Alphabetical sorting hides the point.
3. Start the axis at zero. A bar chart that starts at 80 makes a 5% difference look like a 50% one.
4. Grouped bars for two or three series side by side (this year vs last year by region). Beyond three, use small multiples — one small bar chart per series, same scale.
5. Limit to 12–15 bars. Show the top 10 and group the rest as "Other", with a click-through to the full list.

Example: revenue by product line for 8 lines, this year and last, is a horizontal grouped bar sorted by this year’s value, with last year in grey. One glance shows the leader, the laggard and which one changed places.

Showing change over time: line charts

A line chart shows how a value moves across time and is the right choice once you have more than about 8 points. Below that, bars over time are often clearer because each period reads as a discrete value.

• One line for the metric, one grey line for the comparison (last year, plan). More than four lines becomes spaghetti; split into small multiples.
• Show at least 13 periods for weekly or monthly data so a full year and one extra point are visible.
• Area charts only when the area means something — cumulative revenue against target, for instance. Otherwise the fill adds ink and no information.
• Mark events. A vertical line for a price change or a campaign launch explains a kink in the line better than a caption.

Line axes need not start at zero if the point is the movement rather than the size, but say so in the title ("Conversion rate, last 13 weeks") so readers do not misjudge the scale.

Parts of a whole: stacked bars, and when pie charts are acceptable

Composition questions ("what share…") are where most chart mistakes happen, because the pie chart is the default and it is the wrong tool for most of them.

• Pie or donut: acceptable for 2 or 3 parts where one dominates ("62% of revenue is repeat customers"). Beyond 3 slices, readers cannot compare the angles. Never use two pies to compare periods.
• Stacked bar for the share of several parts across a few categories (revenue by channel for each region). Put the most important part at the bottom so it is measured from the axis.
• 100% stacked bar when the share matters and the total does not (mix of payment methods by month).
• Treemap for a large number of parts with a hierarchy (revenue by category then product), where the reader wants to spot the big and the tiny rather than compare precisely.

If the question is "which part grew?", composition is the wrong frame — a line chart of each part over time answers it directly.

Spread and relationships: histograms, box plots and scatter plots

These two question types come up less often in business dashboards, but when they do, an average hides what matters.

Distribution — how values are spread. A histogram groups order values, delivery times or ticket ages into bands and shows how many fall in each; it reveals the long tail an average conceals (an average delivery of 3 days that is actually 2 days for 90% and 12 days for the rest). A box plot compares spreads across categories — resolution time by support team — in one compact row per team.

Relationship — whether one thing moves with another. A scatter plot puts discount on one axis and margin on the other, one dot per deal; the shape of the cloud answers the question more honestly than a correlation figure. A bubble chart adds a third measure as dot size (deal value) but rarely a fourth as colour — beyond that, it needs a legend nobody reads.

Both types belong on an analysis page, not the KPI page. Put the finding on the KPI page as a number ("18% of deliveries take over 7 days") and link to the chart.

When a table or a single number beats a chart

Two "charts" are not charts and are frequently the best choice.

A table wins when readers need to look up exact values, when there are more than three measures per row (rep, quota, attainment, pipeline, win rate), or when they will sort and scan for exceptions. Make it useful: sort by the column that matters, right-align numbers, highlight cells outside tolerance, and keep it to the rows that fit on a screen with a click-through for the rest.

A KPI tile — one large number with a comparison — wins whenever the question is "what is it right now, and is that good?" Rules: the comparison is mandatory (vs target, vs last period); colour only when off target; an optional sparkline of the last 13 periods underneath; and never more than 7 tiles in a row. The KPI dashboard guide covers how tiles fit into a page.

Charts to avoid, and how to use colour and labels

Avoid:

• 3D anything. Perspective distorts every value.
• Dual-axis charts with two different scales. The crossing point means nothing and readers assume it does. Use two stacked charts instead.
• Gauges and speedometers. A KPI tile with a target says the same in a tenth of the space.
• Radar charts. The area depends on the order of the axes, which is arbitrary.
• Pies with more than 3 slices, and word clouds. Neither can be read precisely.

Colour: one accent for the series that matters, grey for context, and red or amber only for "off target". A rainbow palette makes every series look equally important, which is the same as none. Colour-blind readers are roughly 1 in 12 men; never make red versus green the only signal.

Labels: put the number on the bar or point when there are fewer than about 12; otherwise rely on the axis. Title the chart with the finding ("Returns doubled after the April range change"), not the contents ("Returns by month"). Write units on the axis, and the period and last-updated time somewhere on the page.