Showing whether two things move together.

The only common chart form whose main risk is what the reader concludes rather than what the chart shows.

Drawn as a scatter plot, bubble chart, depending on the data.

People who answered sooner were more satisfied

Rendered from the project Customer survey — Q2, through the same compiler the product uses. Not a picture of one.

What it is for

A relationship chart puts two measures on two axes and shows every observation as a point. It answers whether the two move together, how tightly, and whether the pattern is the same across the whole range or breaks somewhere.

It is also the form most often over-read. Two variables moving together is compatible with one causing the other, the reverse, both being caused by a third thing, or a selection effect in who ended up in the data. A scatter cannot distinguish these, and a chart that cannot distinguish them should not be captioned as though it can.

Reach for it when

Both variables are continuous and measured on the same units of observation
You want to see whether a relationship holds across the whole range or only part of it
Outliers matter individually rather than as a summary
There are enough observations that a pattern could be visible, usually more than thirty

Use something else when

One axis is time, where a trend chart is the direct form
Either variable is categorical
There are so many points that everything overlaps and the plot is a solid block

How this form misleads

Captioning association as cause

The chart shows that two things move together. The caption "faster responses make customers happier" claims a direction the chart cannot supply, and readers remember the caption.

Write the caption as what was observed, and state the plausible alternative explanations beside it. In the worked survey example, people who answered sooner were more satisfied, which is at least as likely to be about who answers quickly.

Overplotting hiding the density

With enough points, the middle of the cloud is solid and the only thing visible is the outer edge. The reader sees the outliers and cannot see where nine-tenths of the data is.

Use transparency, or bin into a heatmap, so that the dense region is readable rather than saturated.

Fitting a straight line through a curved relationship

A linear fit reports a single slope for a relationship that has a threshold or an inflection in it, and the slope describes neither side of the break.

Look at the shape before fitting anything. Where the relationship changes at a point, that point is usually the finding and the line is what hid it.

What this will not draw, whatever you ask for

The agent picks an intent from a closed list and a single compiler renders it. These forms are rejected at the schema layer rather than discouraged in a style guide, which is the difference between a rule and a preference.

3D anything: Perspective makes two equal values look different. There is no case where the third dimension carries data.
Gauges and speedometers: An enormous amount of ink for one number, and the dial implies a range that is usually invented.
Radar charts: Area scales with the square of the values and the shape changes entirely if you reorder the axes.
Word clouds: Size encodes frequency, which is the least interesting thing about a word, and nothing is comparable.
Pie charts above five slices: Angles are hard to compare, and beyond five slices the labels leave the chart anyway.
Per-series arbitrary colour: Colour means something specific in this product. A series that picks its own hue breaks the meaning everywhere else.

Make one from your own data

You do not choose the form. The analysis works out which question your data can answer and picks the chart that answers it, then tells you why.

No account needed to start. You only pay when you like what you see.