Chart selection: how to choose the right chart type for your data, and fix the five charts that mislead

Which chart answers which question?

Start from the question your audience has to answer, then choose the chart whose easiest comparison is that question. William Cleveland and Robert McGill gave the reason in their 1984 paper on graphical perception in the Journal of the American Statistical Association: "Graphs should employ elementary tasks as high in the ordering as possible." In their position-angle experiment, 219 of 4,080 judgments had large errors, 88 percent of those came from angle judgments on pie charts, and the authors calculated that large errors were 7.3 times as frequent for angle judgments as for position judgments on a bar chart.

That experiment had 51 analysed subjects judging what percentage one bar or slice was of the largest, so it does not measure every way people read a chart. It still shows the size of the gap: the same five numbers, read off two charts, produced very different error rates depending only on the encoding.

The map below combines that ranking with the categories of the Financial Times Visual Vocabulary, the chart-selection poster made by the FT Visual Journalism Team for its newsroom.

The question you are answering	FT category	First choice	What the reader compares	Avoid
How did this change over time?	Change over time	Line chart	Position and slope on a shared axis	Cumulative totals, 3D columns
Which is bigger, and by how much?	Magnitude, ranking	Bar chart sorted by value, zero baseline	Length from a shared baseline	Unsorted bars, truncated axis
How are the values spread out?	Distribution	Histogram or box plot	Bar heights across adjacent bins	A single average
Do these two metrics move together?	Correlation	Scatterplot	Position on two axes	Dual axis line chart
Where do people drop out between steps?	Flow	Funnel bar or Sankey	Length of each stage against the first	Pie per stage
What share does each part make up?	Part-to-whole	Stacked bar, or pie with five parts or fewer	Length, or angle for a dominant slice	Pie with many similar slices

Why do some charts read more accurately than others?

Cleveland and McGill listed ten elementary perceptual tasks people perform when they pull a number out of a graph, then ordered them from most to least accurate. The ordering is a hypothesis the paper tests, not a law, and its two experiments found position judgments more accurate than both length and angle judgments.

Rank	Elementary perceptual task	Where you meet it
1	Position along a common scale	Dots or line points on one axis, bar ends on a shared baseline
2	Positions along nonaligned scales	Small multiples with identical axes
3	Length, direction, angle	Stacked segments that do not share a baseline, slopes, pie slices
4	Area	Bubbles, treemaps
5	Volume, curvature	3D shapes, curved areas
6	Shading, color saturation	Heat maps, shaded maps

The practical rule follows directly. Decide which single comparison carries your argument, and put that comparison at rank 1. Everything else on the chart can sit lower, because nobody needs to read it precisely.

What does one dataset look like drawn five ways?

The dataset below is invented for this guide: monthly sign-ups for a B2B invoicing tool by acquisition channel over six months. Total sign-ups grew 16.3 percent, from 1,040 in January to 1,210 in June, while paid social fell 21.1 percent. Each chart below makes a different part of that story easy to see.

Channel	Jan	Feb	Mar	Apr	May	Jun	Six-month total
Organic search	420	450	470	500	540	580	2,960
Paid social	380	410	400	360	330	300	2,180
Referral	150	160	175	190	205	225	1,105
Partner	90	85	95	100	98	105	573
Total	1,040	1,105	1,140	1,150	1,173	1,210	6,818

Drawing 1: a line chart answers "what is each channel doing?"

Months run along the horizontal axis, sign-ups from 0 to 600 on the vertical axis, one line per channel, each labelled at its right end. Organic search climbs steadily from 420 to 580. Paid social peaks at 410 in February and then falls every month to 300. Referral rises from 150 to 225, and partner stays flat around 100. The comparison this chart makes easy is direction: the eye reads four slopes against one axis at once, so the paid social decline is visible in a second even though the total kept growing.

Drawing 2: a sorted bar chart answers "which channel matters most?"

Four horizontal bars, longest at the top, all starting at zero: organic search 2,960, paid social 2,180, referral 1,105, partner 573. The comparison it makes easy is size: organic search delivered about 2.7 times what referral did, and the reader sees that from bar length on a shared baseline. It hides time completely, so it cannot show that paid social is shrinking.

Drawing 3: a 100 percent stacked bar answers "how is the mix shifting?"

Six vertical bars, one per month, each split into the four channels and stretched to 100 percent. Organic search sits at the bottom and grows from 40.4 percent of January sign-ups to 47.9 percent in June. Paid social sits above it and shrinks from 36.5 percent to 24.8 percent. Only the bottom segment and the top segment share a baseline, so those two read accurately; the middle segments float, which drops them to length judgments on nonaligned scales. Put the channel your argument is about at the bottom.

Drawing 4: a slope chart answers "who moved between two dates?"

Two vertical axes, January on the left and June on the right, with each channel drawn as a straight line between its two values. Three lines rise and paid social falls, from 380 to 300. The FT Visual Vocabulary describes the slope chart as good for change "as long as the data can be simplified into 2 or 3 points without missing a key part of story". Here it misses one part: the slope chart cannot show that paid social rose in February before it fell.

Drawing 5: a scatterplot answers "is spend still buying sign-ups?"

This view adds one column the team already had, paid social spend per month: 19,000, 21,000, 22,000, 22,000, 23,000 and 24,000 dollars. Plot spend on the horizontal axis and paid social sign-ups on the vertical axis, one dot per month. The dots run downward to the right: more spend each month, fewer sign-ups. Cost per paid sign-up went from 50 dollars in January to 80 dollars in June. The FT Visual Vocabulary warns that "many readers will assume the relationships you show them to be causal", so title it as what it is: spend rose while sign-ups fell, which points to a saturated audience, not proof that spend caused the drop.

When do you need a distribution or flow chart instead?

A distribution chart answers how values spread out, which no average can. Suppose the same team measured days from sign-up to first invoice sent for 200 new accounts: 64 sent one within a day, 52 within two to three days, 30 within four to seven days, 18 within eight to fourteen days, and 36 never sent one. A histogram with one bar per bucket shows two groups, fast starters and accounts that never start, that a single average would blur into one misleading number.

A flow chart answers where people leave a sequence. For a trial funnel of 1,210 sign-ups, 726 connected a bank account, 484 sent an invoice and 121 paid, a funnel drawn as horizontal bars on one baseline shows each stage against the first: 60 percent, 40 percent and 10 percent. A Sankey diagram earns its complexity only when people can take more than one path, for example when some trials skip the bank connection and send an invoice first.

What are the five charts that mislead, and how do you fix each one?

Each chart below takes true numbers and makes them read wrong, and each fix moves the key comparison back to a shared baseline or puts it in order. Every example uses the dataset above unless it says otherwise.

1. The truncated axis

Before: a column chart of total monthly sign-ups with the vertical axis starting at 1,000 instead of zero. The January column is 40 units tall and the June column is 210 units tall, so June looks 5.25 times as big as January. The real increase is 16.3 percent.

The UK Government Analysis Function guidance on charts states the rule plainly: "Do not break the numerical axis on bar charts", because "breaking the numerical axis distorts these relative proportions." The same page records a real case: an HM Treasury bar chart, criticised by the Office for Statistics Regulation in February 2023, showing inflation falling from 11.1 percent to 10.1 percent on an axis starting at 8 percent. On that axis the second bar is about two thirds the height of the first, while the actual fall is one percentage point, about 9 percent of the starting value.

After: start the column chart at zero. If the change you need to show is small relative to the total, switch to a line chart or a Cleveland dot plot, both of which encode position rather than length, and label the axis so its starting value is obvious.

2. The dual axis

Before: paid social spend as a line on a left axis running from 18,000 to 25,000 dollars, and paid social sign-ups as a line on a right axis running from 280 to 420. Spend rises from low on the chart to high, sign-ups fall from high to low, and the two lines meet exactly in April. A reader concludes something happened in April. Nothing did; the meeting point is a product of where the author set the two scales.

Lisa Charlotte Muth puts it directly in Datawrapper's case against dual axis charts: "The scales of dual axis charts are arbitrary and can therefore (deliberately) mislead readers about the relationship between the two data series."

After: index both series to January equals 100 and plot them on one axis, or plot the ratio between them as a single line.

Month	Spend index	Sign-up index	Cost per paid sign-up
Jan	100.0	100.0	50.0 dollars
Feb	110.5	107.9	51.2 dollars
Mar	115.8	105.3	55.0 dollars
Apr	115.8	94.7	61.1 dollars
May	121.1	86.8	69.7 dollars
Jun	126.3	78.9	80.0 dollars

The indexed version shows the gap between spend and sign-ups widening every month. The cost per sign-up line is usually the better choice for a business audience, because it is the number the decision turns on.

3. The pie with seven slices

Before: a separate survey asked 100 customers who upgraded for their main reason. The answers were needed more seats 19, invoice automation 17, integrations 16, reporting 15, support plan 13, annual discount 11 and other 9. Drawn as a pie, the slices span 68 to 32 degrees, and the four largest differ by 4 points at most, so nobody can rank them by eye.

The UK guidance says a pie should only be used when "there are five categories or fewer", when the categories sum to a meaningful whole, and when there is a dominant category. This pie fails the first and third tests.

After: a horizontal bar chart sorted from 19 down to 9, with the value printed at the end of each bar. If one answer is the point of the slide, colour that bar and leave the rest grey.

4. The cumulative curve

Before: a line of cumulative paid social sign-ups: 380, 790, 1,190, 1,550, 1,880 and 2,180. The line rises every month, as any running total of positive numbers must, and looks like success. Monthly paid social sign-ups fell 21.1 percent over the same period; the only trace of that in the cumulative line is a slight flattening of the slope that few readers notice.

After: plot the monthly values, which is drawing 1 above, or the change from the previous month. Keep cumulative charts for one job they do well: progress toward a fixed target, with the target drawn as a horizontal line.

5. The unsorted bars

Before: trial-to-paid conversion by country in alphabetical order: Brazil 11.2 percent, France 14.8, Germany 16.1, India 8.4, Netherlands 15.3, Portugal 12.6, Spain 13.9, United Kingdom 17.0. To find the best and worst markets the reader scans back and forth across eight bars.

The UK guidance says to "Always rank bars by value, unless there is a natural order, for example, age or time." The FT Visual Vocabulary adds that standard bar charts display ranks "much more easily when sorted into order".

After: sort from the United Kingdom at 17.0 percent down to India at 8.4 percent. The top three and the outlier at the bottom now read in one glance, and the question the meeting should discuss, why India converts at half the UK rate, is on the slide without anyone saying it.

What is the strongest argument against these rules?

Strict rules can make a chart less honest. A zero baseline on a line chart of a metric that moves between 94 and 97 percent flattens every change into a straight line, and the UK guidance shows its own employment rate chart with a non-zero axis, clearly labelled, because the changes are the point. Cleveland and McGill also say accuracy is not everything: "We do not argue that this accuracy of quantitative extraction is the only aspect of a graph for which one might want to develop a theory, but it is an important one."

So the rules split by encoding. Bars encode length, so their baseline stays at zero. Lines and dots encode position, so their axis can start where the data starts, as long as the label says so. A pie is acceptable when one slice dominates and there are only two or three parts, because then the only judgment the reader makes is "more than half or not".

How do you check a chart before it goes in front of stakeholders?

Write the sentence the chart should prove before you draw it; the UK guidance suggests imagining you are describing the chart over the phone. Then run three checks:

• The comparison check. Name the one comparison your sentence depends on and confirm it sits on a shared baseline or a common axis.
• The scale check. Bars start at zero, dual axes are gone, and any line chart that starts above zero says so on the axis.
• The order check. Bars are sorted by value unless the categories have a natural order, and the category your argument is about is the one coloured.

A chart that passes all three is ready for a title that states the finding, which is where how to present data to executives picks up. For the difference between showing a number and stating an insight, see data storytelling; for deciding which metric deserves a chart at all, see how to measure feature success. When the chart goes into a decision deck, AI for product presentations covers the slide around it.

Builders Camp's Turning Data Into Storytelling bootcamp runs over 1 week with 2 live sessions and 9 self-paced microlessons, and chart selection and visual clarity is one of the six topics on its public syllabus.

See the Turning Data Into Storytelling bootcamp

Take the last chart you presented and redraw it with the comparison that mattered moved to rank 1 of Cleveland and McGill's list. If the redrawn version tells a different story from the original, the original was the one making the argument, not the data.