Chart chooser for real life

Someone asks for “a nice chart of the funnel.” You open the template gallery. There is a 3D pie, a stacked area that looks like a mountain range, and a radar chart that somehow made it into the product. You pick the one that looks busiest, because busy reads as thorough. Two minutes later the room argues about the colors instead of the conversion drop between demo and paid.

This is Part 2 of Charts that make sense. Part 1 covered purpose: explore, explain, persuade. Now we pick a shape that matches the question. Chart choice is not fashion. It is matching a relationship in the data to a mark people can read under fluorescent lights at 8:55 a.m.

What you’ll learn

Four everyday chart jobs: comparison, trend, composition, distribution
A simple chooser you can run before you open Sheets, Tableau, or Power BI
When tables beat charts
A worked example that maps one sales dataset to four views
A never-do-this list for pies, 3D, dual axes, and rainbow overload

Start with the relationship, not the template

Gallery tools sell templates. Your brain needs relationships. The Financial Times Visual Vocabulary and many teaching charts from Stephen Few’s work organize graphics by what you are trying to show: magnitude, change over time, part-to-whole, distribution, correlation, and so on. You do not need the full museum map on day one. You need four relationships that cover most workplace decks.

Ask: what is the main relationship I want someone to see in under five seconds? Then pick the mark that makes that relationship cheap to see. If two relationships both matter, you may need two charts. One overloaded chart is not thrift. It is confusion with a legend.

Rule of thumb: Choose the chart that makes the wrong conclusion hard, not the chart that makes the right conclusion look fancy.

1. Comparison: who or what is bigger?

Comparison answers ranking and magnitude questions. Which region sold more? Which product line is slowest? Which campaign spent more than it returned?

Example chart for comparison:

Sorted category bars answer “who is bigger?”

Default marks: horizontal or vertical bars. Horizontal bars shine when category names are long. Sort intentionally (by value for ranking, by a natural order only when the order is the story).

Good for: categories with one primary measure. A few series side by side if the series are few and named clearly.

Weak for: long time series (use a line), subtle part-to-whole shifts (consider stacked bars carefully or a table of shares), and tiny differences across dozens of categories (filter, group, or use a table).

Workplace example: “Which three support queues need overtime this weekend?” A sorted bar of open tickets by queue is almost always better than a pie of “queue mix,” because the decision is ranking, not pie slices.

2. Trend: how did this change over time?

Trend answers change, seasonality, and break points. Did orders recover after the outage? Is the weekly active user line flat while marketing spend rises?

Example chart for trend:

Line over ordered time for change, not for unordered categories.

Default marks: lines for continuous time, connected dots when points are sparse. Bars can work for a few periods, but lines encode change more clearly when the path matters.

Good for: one primary series, or a few series that do not braid into spaghetti. Annotations for events (launch, outage, price change) beat rainbow legends.

Weak for: many categories over time (prefer small multiples or a ranked snapshot), and part-to-whole that must sum exactly every period (stacked areas can mislead when the eye tracks tops instead of segments).

If you already built time aggregates in SQL or pandas, keep the grain honest. A daily line with weekend zeros can look like a heart monitor. That is a data shape issue as much as a chart issue. Foundations and quality work still apply: see Analytics foundations and data quality.

3. Composition: what makes up the whole?

Composition answers part-to-whole questions. What share of revenue is each plan? What mix of ticket types arrived this week?

Example chart for composition (stacked bars; keep category count modest):

Stacked bars for mix over a few periods.

Default marks: stacked bars for comparing composition across a few periods or categories; a simple bar of shares when you care about one period; a table when exact percents matter more than a picture. Pies are optional and fragile: best with two to four large slices and clear labels, never as a default for twelve product lines.

Good for: a small number of parts, stable definitions, and audiences who already know the total.

Weak for: many tiny slices, compositions that change definition mid-series, and ranking tasks that people keep misreading as “who is biggest” inside a pie.

If the real question is ranking, do not force composition. “East is 41% of tickets” is composition. “East has the most tickets” is comparison. They are related. They are not the same slide.

4. Distribution: how are values spread?

Distribution answers spread, skew, and outliers. Are deal sizes clustered, or does one whale deal invent the average? How long do tickets stay open for most customers?

Example chart for distribution:

Histogram shows spread that a single average hides.

Default marks: histogram for one continuous measure, box plot or violin when comparing spreads across groups, strip or beeswarm when you need to see individuals without pretending they are a single average.

Good for: challenging average-only stories, quality checks, and capacity planning where the tail matters.

Weak for: executive slides that only need a ranking, and tiny samples where a histogram looks scientific while hiding “n = 7.”

Distributions are underrated in decks. Many bad decisions start with “the average is fine” while half the customers wait twice as long. If you only show a mean bar, you are not showing the distribution job. You are showing a summary that may lie about risk.

A useful workplace phrase is “show me the middle and the tail.” The middle can be a median. The tail can be the 90th percentile, a long-wait count, or a simple histogram. Customer support, logistics, and sales cycles all live in the tail more than leaders admit. If your chart type cannot show spread, do not pretend a single bar answered the risk question.

Another distribution trap is mixing apples in one histogram: ticket ages for self-serve and enterprise together, or deal sizes across products with totally different price points. When groups differ that much, compare distributions with small multiples or separate panels. One combined histogram can invent a “bimodal mystery” that is only two products sitting in the same plot.

The 30-second chooser

If the question sounds like…	Relationship	Start with
Which is biggest / smallest?	Comparison	Sorted bars
How did this change?	Trend	Line over time
What share of the total?	Composition	Stacked bar or share bars
How spread out are values?	Distribution	Histogram or box plot
Do two measures move together?	Relationship	Scatter (use carefully)
Exact values for lookup?	Reference	Table

Scatter plots earn a seat when correlation or clustering is the claim. They fail when people treat cloud shape as proof of causation. For most Monday decks, the four big jobs above cover the work.

When a table is the better “chart”

Analysts sometimes treat tables as the thing you graduate away from. That is backwards. Tables win when:

Exact values matter more than pattern
There are few rows and many measures
People need to look up a single cell (“What was West in W4?”)
You would need a legend novel to explain the graphic

A clean table with one highlight can beat a decorative radial chart every day of the week. Charting is optional. Clarity is not.

Worked example: one dataset, four questions

Here is a small product sales extract for one quarter. Units are order counts. Dollars are revenue.

Product	Jan orders	Feb orders	Mar orders	Q1 revenue
Starter	120	130	125	$48,000
Pro	80	95	110	$132,000
Enterprise	12	15	18	$210,000
Add-ons	200	210	230	$41,000

Question A: comparison

“Which product drove the most revenue in Q1?” Sorted horizontal bars of Q1 revenue. Enterprise wins. Order counts would mislead, because Add-ons are frequent and cheap. The chart type follows the measure that matches the decision.

Question B: trend

“Is Pro gaining momentum in orders?” Lines for Pro orders Jan through Mar, maybe with Starter as a quiet comparison. A pie of March mix does not answer momentum.

Question C: composition

“What is the revenue mix?” Share of Q1 revenue as a simple bar of percents, or a stacked bar if you also show last quarter. A four-slice pie can work if labeled with values, but bars make Enterprise’s dominance easier to compare later when you add a fifth product.

Question D: distribution

This summary table cannot show distribution of deal sizes inside Enterprise. If the real risk is that three giant deals created the average, you need the deal-level extract and a histogram or box plot. That is the point: the chooser also tells you when the grain of the table is wrong for the question. Go back to SQL or the warehouse, not to a fancier chart of the wrong grain. The SQL series and Python series are for that handoff.

# Pseudo: map question -> view (not a full plotting tutorial)
question = "which product drove revenue"
if question.startswith("which") or "rank" in question:
    chart = "sorted_bar"
elif "over time" in question or "momentum" in question:
    chart = "line"
elif "share" in question or "mix" in question:
    chart = "share_bar_or_careful_stack"
elif "spread" in question or "typical deal" in question:
    chart = "histogram_or_box"
else:
    chart = "table_until_sure"
# If grain is wrong for chart, stop and re-aggregate upstream.

Never-do-this (and what to do instead)

Never	Why it fails	Do this instead
3D bars, pies, or exploded slices	Perspective distorts magnitude	Flat 2D bars or lines
Pie with 8+ slices	People cannot compare angles	Sorted bars or a table
Rainbow categorical colors by default	Legend hunting, print fails	Few hues, labels on marks
Dual axis for unrelated metrics	Invented “relationships”	Two charts or indexed series
Stacked area for many series	Middle layers are unreadable	Lines, small multiples, or ranking
Map when region names would do	Pretty geography, weak ranking	Bars sorted by value
Radar / spider for ordinary KPIs	Hard to read, easy to overclaim	Bars or a simple table

These are not aesthetic preferences. They are readability preferences under real meeting conditions: low projector contrast, people in the back, and five minutes on the agenda. If a chart only works when you explain it for three minutes, the chart is still doing explore work in an explain slot.

Also watch for “chart type drift” across a series of weekly slides. Week one uses bars because ranking mattered. Week four silently becomes a stacked area because someone liked the look, and now the room is back to debating colors. Keep the type stable when the question is stable. Change the type when the question changes, and say so in one line of the title so people are not relearning the visual language every Monday.

If you hand off charts from SQL results, keep the aggregation close to the question. A perfect chooser cannot save a grain that is wrong. Build the table for the relationship you need, then pick the mark. That order matches how solid analytics work already runs in the SQL series: query for the decision, not for a random dump that you hope a chart will magically clarify.

Common mistakes

Picking the template that matches brand colors, not the question.
Using composition charts for ranking decisions.
Showing trends with unordered categories on the x-axis. Time belongs in order. Priority labels do not.
Spaghetti lines: eight series, one plot, no small multiples.
Average-only bars when the distribution is the risk.
Charting before checking definitions. A perfect line of the wrong metric is still wrong. Quality first when totals disagree.
Assuming interactive hover will save a dense static export. PDF and screenshot kill hover.

How to practice this week

Take three charts from a recent deck. Label each comparison, trend, composition, or distribution.
For each, write a better question the chart should answer. If the type does not match, redesign it in ten minutes.
Replace one pie with sorted bars and show both to a colleague. Ask which ranking they can read faster.
Find one place where a table would beat a chart and ship the table without apology.
Keep a personal chooser sticky note with the four jobs until muscle memory arrives.

More learning paths live on Learn. Next up: axes, scales, and honest baselines, because a good chart type with a dishonest scale still misleads.

Quick recap

Match relationship to mark: comparison, trend, composition, distribution.
Start from the question, not the template gallery.
Tables are valid. 3D and pie overload usually are not.
Wrong grain means fix the data step, not the chart skin.
One chart, one primary relationship.

Sources

Financial Times, Visual Vocabulary: https://github.com/ft-interactive/chart-doctor/tree/master/visual-vocabulary
Stephen Few, Perceptual Edge: https://www.perceptualedge.com/
Alberto Cairo, functional visualization teaching: https://www.albertocairo.com/
Datawrapper Academy, chart type guidance: https://www.datawrapper.de/charts
From Data to Viz (decision tree for chart types): https://www.data-to-viz.com/
matplotlib gallery (implementation examples when you plot in Python): https://matplotlib.org/stable/gallery/index.html
Analytics Made Simple, Charts that make sense Part 1 context via Learn: https://analyticsmadesimple.com/learn/