Twenty-five countries need comparing on one slide, and there is a spreadsheet open, four chart types tried in four browser tabs, and eight minutes before the meeting starts. Which chart should I use? sits at the centre of almost every slide that goes wrong, and the reason is rarely taste. Most bad charts fail for a handful of structural reasons, the wrong chart type for the shape of the data, long before anyone gets to colour or font. Knowing how to choose the right chart is mostly knowing what the chart is for.
The fix is to stop starting from the chart and start from the purpose. What does this chart actually need to say? A comparison across categories, a change over time, a piece of a whole, a spread of values, a relationship between two numbers, or a ranking: pick one, and the field of reasonable options narrows fast. This guide walks through each purpose, with chart type examples, a short table for reference, and a chooser tool below that does the same job interactively.
A short guide to how to choose the right chart
The short version: name the purpose first, then pick from the small set of charts built for it. Comparing categories points to bars. Change over time points to lines, unless there are only two or three points, which points to columns or a slope chart. Parts of a whole point to a pie only under five or six slices, and to a stacked bar or treemap above that. A distribution points to a histogram or box plot. A relationship between two numbers points to a scatter plot. A ranking points to sorted bars or a dot plot, and a ranking that moves over time points to a bump chart. Everything below expands on why, with the traps that undo each one.
Why position beats angles and areas
Before the chart families, it helps to know why some chart types keep winning over others that look more interesting. In 1984, statisticians William Cleveland and Robert McGill published a study in the Journal of the American Statistical Association, “Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods,” testing how accurately people judge different visual properties. Position along a shared scale, the kind a bar chart or dot plot uses, came out as the most accurately judged property. Angles, the kind a pie chart uses, came out noticeably less accurate, and area judgements (the basis of a bubble chart or a 3D pie) fared worse still.
That is not an argument against pie charts entirely. It is an argument for knowing what a chart type costs in accuracy before choosing it for anything other than a very small, very simple comparison. A pie with three slices, one obviously dominant, still reads fine. A pie with fourteen roughly equal ones is asking the reader to do the hardest kind of visual judgement Cleveland and McGill tested, on the least forgiving chart for it.
The Financial Times’ own Visual Vocabulary, built by its visual journalism team and inspired by Jon Schwabish and Severino Ribecca’s earlier Graphic Continuum, organises the whole field of chart types the same purpose-first way, in nine groups: deviation, correlation, ranking, distribution, change over time, part-to-whole, magnitude, spatial and flow, rather than as a gallery sorted by how each chart looks. A marketing professor, Andrew Abela, built a similar decision tree around 2009 for consultants choosing between comparison, composition, distribution and relationship, and it is still handed around in workshops today, in PDF form, nearly two decades on. The tool below this guide, and the dedicated chart chooser tool, both work from that same logic: ask what the chart needs to prove, then narrow the list.
Comparison: bars, columns and dot plots
Comparing values across categories is the single most common chart job, and bar and column charts exist because they are the most accurate way to do it. A column chart with up to about seven categories, sorted by value and starting at zero, is close to unbeatable: the eye compares lengths, which Cleveland and McGill’s research ranked as the most reliable visual judgement people make.
Two or three series per category call for grouped bars, placed side by side inside each category so the within-category comparison stays direct. Four or more series turn a grouped bar chart into a comb that is hard to scan, and small multiples (one small chart per series, sharing the same scale) usually work better at that point. Once the category count climbs past twenty, switch from bars to a sorted dot plot; a dot uses far less ink than a bar and does not need to start at zero, so the scale can zoom into the range where the real differences actually sit.
The one rule that survives every variation here: start the value axis at zero. A bar’s entire argument is its length, and a value axis cut at, say, ninety instead of zero can make a fifteen percent difference look four times larger than it is. It is the single fastest way to turn an honest comparison into a misleading one, discussed in more detail in the lesson on how to spot a misleading graph.
Change over time: lines, columns and slope charts
A line chart connects points in time order, so trend, direction and turning points read in one sweep. It works well with up to about four series and enough points (three to twelve is comfortable, more than that fine too) to make a slope worth drawing. Fewer than three points and a line starts implying a smoothness the data never had; a column chart or a slope chart communicates two or three points more honestly.
A slope chart in particular deserves more use than it gets. Two vertical axes, one per date, joined by a line per item: the angle of each line shows the size and direction of the change, and it is a genuinely clear way to show “before and after” across several items at once, something a line chart with only two data points struggles to make visually interesting.
Beyond four series, a line chart becomes a tangle of crossing colours nobody can trace reliably. Small multiples, the same chart repeated once per series with each panel sharing identical axes, solve that by giving each line its own space instead of forcing them to share one plot area.
Parts of a whole: pies, stacked bars and treemaps
A pie or donut chart earns its keep with five parts or fewer, when one slice is meant to look dominant or two slices close in size are the actual point. Past five or six slices, the angles crowd together and a reader cannot reliably tell whether one slice is bigger than another without reading the printed number, which defeats the purpose of drawing a picture in the first place. At that point, a 100% stacked bar chart or a sorted horizontal bar chart with percentages labelled communicates the same composition with far less visual guesswork.
Nested hierarchies, parts within parts like a budget broken into departments and then line items, suit a treemap, where nested rectangles use area to show size at two levels simultaneously. Several separate wholes compared side by side, such as market share across five regions, suit small multiples of 100% stacked bars rather than a row of pies, which the misleading-graph lesson names as one of the most reliable ways to confuse a reader with an honest chart type used the wrong way.
A running total built from additions and subtractions, starting revenue, several gains and losses, ending revenue, is its own case: a waterfall chart, with floating bars anchored to a shared baseline at the start and the end, answers “what moved the number” directly, something no static pie or stacked bar attempts.
Distribution: histograms, box plots and strip plots
A distribution question (how spread out are these values, where do most of them sit, is there a long tail) cannot be answered by a single average sitting alone in a bar chart. An average hides exactly the information a distribution chart exists to reveal.
For one group with fewer than about fifty values, a strip plot showing every individual point on a shared scale keeps the full picture intact, with nothing summarised away. Past fifty values, a histogram, which counts values into equal-width bins, shows the shape: where the bulk sits, whether it skews toward one end, whether there are two separate peaks hiding inside what looked like one group. Comparing that shape across several groups at once calls for a box plot, which compresses each group into its median and middle half so several distributions can line up side by side without the panel-per-group cost small multiples would carry.
Relationship: scatter plots and their heavier cousins
Two variables, one relationship to test: a scatter plot, one dot per item, placed by its two values, is the standard tool, and for good reason: position on both axes is exactly the kind of judgement Cleveland and McGill’s research found people make most accurately. Patterns, clusters and outliers become visible immediately, without a formula or a summary statistic standing between the reader and the data.
A third variable carried as the size of each point turns a scatter plot into a bubble chart, useful when one of the three values is genuinely a volume (population, revenue, headcount), but weaker than the scatter’s two axes, since people read size less precisely than position. Four or more variables at once call for a correlation heatmap, a grid of every pair shaded by how strongly they move together, useful as an overview before picking one pair to examine properly in a scatter plot. And once a scatter plot’s dots number in the tens of thousands, the chart turns into a solid blob; a binned heatmap, counting points into a grid of cells, shows density where individual dots no longer can.
Ranking: sorted bars, dot plots and bump charts
A ranking at one point in time is a comparison chart wearing a different hat: sorted bars for up to about twenty items, a sorted dot plot beyond that. The sorting itself does most of the communicating: an unsorted or alphabetical bar chart forces the reader to scan the whole thing just to find the leader, which is precisely the kind of avoidable friction a ranking chart should remove.
A ranking that changes across several dates is a different problem entirely, and a line chart of the raw values usually answers the wrong question, since rank and value are not the same thing. A bump chart plots rank directly, one line per item across the dates, so every overtaking move becomes a visible crossing point. It is built for exactly the kind of “who’s in the top ten this year versus last year” question a league table or a market-share leaderboard raises constantly.
A note on data that doesn’t fit any of these six purposes
Not every dataset sorts cleanly into comparison, time, parts, distribution, relationship or ranking, and forcing one onto data that resists it is its own kind of chart-selection mistake. Text that is mostly quotes and short answers, for instance, is often better served by a table or a short list than by any chart at all. Turning four survey answers into a bar chart just because the source happened to be numeric adds a decoding step the reader did not need. When a dataset genuinely straddles two purposes, such as a ranking that also needs to show the size of the gap between items, it is usually better to pick the purpose the audience cares about more and let the chart answer that question clearly, rather than building one chart that tries to answer two questions at once and does neither well. If it still isn’t obvious which chart should I use for a specific dataset, working backward from the sentence the slide is meant to prove, not the columns in the spreadsheet, almost always narrows it back down to one of the six.
Chart type examples worth studying
A quick reference, gathered from the framework above, of which charts suit which goal and which charts tend to work against the reader instead:
| Goal | Charts that work | Charts that usually mislead or slow the reader |
|---|---|---|
| Compare categories | Sorted bar/column, grouped bars, dot plot (many items) | 3D bars, a cut value axis, pies for unrelated categories |
| Change over time | Line, column (few points), slope chart, small multiples | One pie per period, more than four overlapping lines |
| Parts of a whole | Pie (≤5 parts), 100% stacked bar, treemap, waterfall | 3D or exploded pies, 6+ slices, rows of pies to compare |
| Distribution | Histogram, box plot, strip plot | A single average in a bar, a line chart across unordered bins |
| Relationship | Scatter plot, bubble chart, binned heatmap | Dual-axis lines presented as proof of a link, 3D scatter |
| Ranking | Sorted bars, dot plot, bump chart (over time) | Unsorted or alphabetical bars, pies, radar charts |
Not every document reaches for a chart at all, and that is sometimes the more disciplined choice. The European Athletics Championships Statistics Handbook, compiled edition after edition by the statistics firm Tilastopaja Oy, pairs its results tables with an athlete index and a country index rather than turning ninety years of championship results into a chart, because the actual reader need is precise lookup, name by name, not a visual comparison. A results table sorted correctly is, in its own way, a chart choice too.
Which chart should I use? Practising the choice
Reading the framework above is one thing; applying it under time pressure with a real spreadsheet open is another. The chart chooser walks through two or three short questions about the data’s purpose and returns a recommended chart with the reasoning and the traps spelled out, working from the same purpose-first logic covered here. The game below flips the exercise around: twelve real questions, each already paired with a chart type, and the job is to keep the pairing when it holds up and toss it when it doesn’t. That is a faster way to build the same instinct than reading the rules a second time. Either tool answers the same underlying question, which chart should I use, faster than scrolling back through this guide from the top every time.
What tends to go wrong even after the right chart is picked
Choosing the right chart type solves the biggest failure mode, but not the only one. A correctly chosen bar chart with a legend detached three inches away, or a correctly chosen line chart with gridlines darker than the lines themselves, still asks more of a reader than it should. Those problems belong to slide and chart craftsmanship rather than chart selection, and the misleading graph and slide design checklist lessons cover them directly. Chart selection is the first decision, not the last one, and it is the one that decides whether every later fix is even possible.
A last practical note on how to choose the right chart under real deadline pressure: write the sentence the slide is supposed to prove before opening any chart menu at all. “Region three grew fastest” points straight at a sorted bar chart. “Churn dropped after the March change” points straight at a line or a labelled column pair. The sentence almost always names its own chart type; the tool below just makes that mapping explicit for the cases that feel less obvious.