Fix this chart · Edit Distance

A chart is an argument. Most bad charts are not lies about the numbers, they are lies about the drawing: where the axis starts, which scale sits on the right, how many slices, how much depth. Below is one small made-up dataset drawn six ways. Each one tells a different story. Each one has a toggle that redraws it honestly.

Truncated y-axis

Orders index, axis starts at 99.4. Illustrative data.

Two made-up series over 24 months: an orders index (about 103 to about 100) and monthly sign-ups (a peak at month 11, then a decline). The shapes are the point, not the values.

One dataset, six stories

Two made-up series over 24 months. An orders index that drifts from about 103 to about 100, a 3% decline. Drawn here with the y-axis starting just under 100, it looks like a cliff. Hit "Fix it" to start the axis at zero and watch the cliff turn into a shelf.

Two axes, one invented correlation

Put sign-ups on a second axis and the two lines look like they move together. They do not. The right-hand scale was chosen to make them touch. The fix is two charts, each with its own zero.

Nine slices

Which is bigger, Referral or Email? A pie asks you to compare angles, and past six slices nobody can. Sorted bars answer the question before you finish reading the labels.

Rainbow

Rainbow palettes put two greens next to each other and a yellow that vanishes on white. A fixed-order series palette gives each channel a colour it keeps in every chart on the page, and the legend follows the same order.

Cumulative always goes up

A running total cannot fall. Sign-ups peaked at month 11 and have halved since, and the cumulative line shows a healthy climb the whole way. Per-month bars show the decline.

Fake depth

The 3D bars are the same eight quarterly numbers. Perspective moves every bar top away from the gridline that should read it. Flat bars, values on top, done.

All six

Every chart here is the same two series. Flip any of them. The question to ask of any chart you meet is the same one each toggle answers: what would this look like drawn honestly?

The one question

Every trick above is defeated by the same question: what would this look like drawn honestly?

Not “is this data right”. The data in all six charts is identical and it is right, for whatever a made-up dataset is worth. The question is about the drawing. If you can picture the chart with a zero baseline, one axis, sorted bars, a fixed palette, per-period values, and no perspective, and the picture changes the story, the story was in the drawing and not in the data.

I ask it of my own charts first. It is much easier to catch a truncated axis when you are the one who cropped it.

~3%

actual decline in the orders index

illustrative series, month 1 to month 24

~50%

drop in monthly sign-ups since the peak

illustrative series, month 11 to month 24

ways the same two series were drawn

all in the piece above

Six sins, one toggle each

The truncated axis is the most common one because it is the default in a lot of tools. Auto-scaling picks the tightest range that fits the data, and a 3% drift becomes a cliff. Zero is not always the right baseline (temperature, index values against 100), but if you leave it out you should say so on the chart.

Dual axes are the sin I see most from smart people. Two series, two scales, one plot, and the right-hand scale gets chosen until the lines cross where you want them to. There is no correct way to align two arbitrary scales, which means the apparent correlation is a choice. Two stacked charts with a shared x-axis take the same space and make no claim.

A nine-slice pie is not a chart, it is a legend with decoration. The reader cannot compare angles past about six slices, and the colours run out at eight. Sorted bars answer the only question a part-to-whole chart is for: which is biggest, and by how much.

Rainbow palettes look lively and read terribly. Adjacent hues collide, yellow disappears on a light surface, and the colours carry no memory from one chart to the next. A fixed-order series palette gives each entity a colour it keeps everywhere. Filter out a series and the survivors keep their hue.

Cumulative lines are the sin people commit on purpose. A running total cannot go down, so a metric in decline still produces a chart that rises. In the piece above, sign-ups peak at month 11 and roughly halve by month 24. The cumulative line climbs the whole way. Per-period bars show what is happening now.

3D bars are the easy one to mock and still the one that shows up in board decks. Perspective moves every bar top off the gridline that should read it, and the front and back faces suggest two different values. Flat bars with the value on top cost nothing and read in one glance.

The checklist

I keep this short so I actually use it. Before a chart ships, or before I believe one:

Where does the y-axis start, and if not at zero, does the chart say so?
Is there a second y-axis? If yes, why not two charts?
Is it a pie? If yes, does it have more than four slices?
Do the colours mean anything, and do they mean the same thing on the next chart?
Is it cumulative? If yes, what does the per-period version look like?
Is there depth, shadow, or perspective doing work that a value label could do?

If a chart fails two of these it is not a chart, it is a pitch. Redraw it before you decide anything with it.

What to do

Take the last chart you sent someone and draw it the honest way. If the story changes, send a correction.
Set your charting defaults to a zero baseline, a fixed series palette, and no 3D. Most tools let you.
When a chart in a deck moves you, ask the presenter for the version with a zero axis and one scale. Watch what happens.
Keep the checklist near the chart tool, not in a wiki nobody opens.