Which Chart Should You Use? A Decision Guide by Data Type

Most bad charts aren’t badly made — they’re the wrong chart, made well. Someone put real effort into formatting a pie chart that should have been a bar chart, or polishing a dual-axis monstrosity that should have been two separate charts, and no amount of polish fixes a wrong-tool decision.

The good news: chart choice isn’t a matter of taste. Your data has a shape — it’s comparing, trending, distributing, flowing, or relating — and each shape has one or two charts that serve it well and a dozen that don’t. This guide gives you the decision in table form first, then the reasoning, then the list of charts you should stop using regardless of what your data looks like.

One principle governs everything here: choose the chart for your reader, not for your data. The question is never “what can I plot?” It’s “what should someone understand in the first three seconds of looking at this?”

The decision matrix

Find what your data is doing in the left column. The verdict column is our recommendation; the alternatives are legitimate; everything else is noise.

Your data is…	Use this	Acceptable alternative	Build it
Comparing values across categories	Bar chart (horizontal if labels are long)	Column chart	Native in Excel/Sheets
Showing change over time	Line chart	Column chart (few periods)	Line graph tutorial
Comparing categories and time together	Clustered column, or small multiples	Line chart, one line per category	Native
Explaining how a total was built up or eroded	Waterfall chart	Stacked column	Excel · Google Sheets
Showing part-to-whole composition	100% stacked column (over time) or waffle chart (single snapshot)	Pie — max 4 slices, sorted	Waffle chart tutorial
Finding the vital few causes (80/20)	Pareto chart	Sorted bar	Pareto tutorial
Tracking KPIs against targets	Bullet chart	Actual-vs-target columns	Bullet chart tutorial
Two different units or scales together	Combo chart	Two stacked charts sharing an x-axis	Combo chart · Secondary axis
Showing intensity across two dimensions	Heat map	Conditional-formatted table	Heat map tutorial
Showing a distribution	Histogram / bell curve	Box plot (via Python in Copilot)	Bell curve tutorial
Stage-by-stage drop-off	Funnel chart	Bar chart per stage	Funnel tutorial
Before-vs-after for many items	Dumbbell chart	Paired bars	Dumbbell tutorial
Sensitivity of an outcome to inputs	Tornado chart	Sorted bar of impacts	Tornado tutorial
Flows between states or categories	Sankey diagram	Stacked columns per stage	Sankey tutorial
Relationship between two variables	Scatter plot (+ trendline)	—	Native
Events and milestones in sequence	Timeline chart	—	Timeline tutorial
A single KPI as a dial	Bullet chart (yes, really)	Gauge, if you must	Gauge tutorial

Two structural notes before the verdicts. First, if your chart needs to update as data grows, build it on a Table or dynamic arrays from the start — our dynamic charts guide covers the modern method. Second, if your data needs summarizing before it can be charted at all (raw transactions, not totals), the right move is a PivotChart, and the complete PivotCharts guide is where to go.

The verdicts, scenario by scenario

Comparing categories: bars, and go horizontal more often

For “which region sold most,” a bar chart wins and it isn’t close. Use horizontal bars whenever category labels exceed a word or two — rotated or truncated labels on column charts are the most common self-inflicted readability wound in business reporting. Sort by value, not alphabetically, unless the category order carries meaning. An unsorted bar chart makes your reader do the ranking work the chart exists to do.

Change over time: lines, with one honest exception

Time on the x-axis, values as a line. The exception: with only a handful of periods — quarterly results for one year, say — columns often read better because they emphasize each period as a discrete result rather than implying a continuous trend through four points. Beyond roughly eight periods, columns turn into a picket fence; switch to a line.

Composition: the pie chart gets a probation notice, not a pardon

The honest verdict on pies: humans judge angles poorly and lengths well, which is why a sorted bar chart nearly always communicates composition better. But a pie with at most four slices, sorted largest-first, with the message in one dominant slice (“one product is 61% of revenue”) is legitimate — the sin isn’t the pie, it’s the eleven-slice pie with a legend like a phone book. For a single part-to-whole snapshot with more precision, a waffle chart beats both: ten-by-ten grid, each square a percent, no angle-guessing. For composition over time, use a 100% stacked column and skip the pie debate entirely.

Totals that build or erode: the waterfall earns its popularity

When the story is “we started at X, these things added, these things subtracted, we ended at Y” — profit bridges, headcount movements, cash flows — nothing beats a waterfall chart. It’s the one finance-fashionable chart that deserves its status. Modern Excel and Google Sheets both have native waterfall types now; if you’re following a tutorial that starts with hidden helper columns, you’re reading something written before 2020.

Targets and KPIs: bullet charts, and the case against gauges

Here’s a position we’ll defend all day: the gauge chart is a waste of dashboard space and the bullet chart should replace it almost everywhere. A gauge spends a large circle communicating one number against one threshold. A bullet chart communicates the actual, the target, and qualitative bands (poor/fair/good) in a slim horizontal strip — you can stack eight of them in the space one gauge occupies. Gauges survive because they look like car dashboards and executives find them familiar. Familiarity is not a data-visualization argument. (If the boss insists, here’s the gauge tutorial — we’re opinionated, not unemployed.)

Two scales on one chart: allowed, with a warning label

Revenue in dollars and margin in percent on the same chart requires a combo chart with a secondary axis, and it’s a legitimate tool — that’s why we teach it. But dual axes are also the easiest way to manufacture a correlation that doesn’t exist, because the reader’s eye compares line positions that the axis scales, not the data, determined. Rules: only when units genuinely differ, always with unmistakable series-to-axis labeling, never two similar lines. Our secondary axis guide includes the full “when you shouldn’t” argument.

Diagnosis charts: Pareto, tornado, funnel, dumbbell

These four are underused because people don’t know their names, and each is the unambiguous winner for its scenario. Ranking causes by cumulative impact — complaint sources, defect types — is a Pareto chart. Showing which assumption moves a model most is a tornado chart. Stage-by-stage conversion drop-off is a funnel. Before-vs-after across many items — salaries after a review cycle, scores after training — is a dumbbell chart, which beats paired bars because the connecting line makes each gap the visual object.

Density, distribution, and flow: the specialist bench

A heat map is the right call when both axes are categories and the value is intensity — sales by weekday and hour, risk by likelihood and impact. It’s a table your reader can absorb in a glance. Distributions (“what’s typical, what’s an outlier?”) want a histogram or bell curve, not averages — an average hides exactly what a distribution question is asking about. And flows between states — energy in/out, traffic between channels, budget allocation paths — are Sankey territory; we’ll be straight with you that Excel makes Sankeys grudgingly, and the tutorial includes an honest verdict on when to use a free external tool instead.

For events in sequence — project milestones, launch plans — use a timeline chart. If what you actually need is task bars with durations and dependencies, that’s a Gantt chart, which is a project-management tool rather than a data chart; a dedicated resource like gantt-chart.co.uk covers it properly.

When several categories trend together: small multiples over spaghetti

Six product lines over 24 months on one line chart produces spaghetti — intersecting lines nobody can follow. The professional answer is small multiples: the same simple chart repeated once per category, identical axes, arranged in a grid. Each panel is trivially readable, and the shared axes make cross-panel comparison honest. Excel has no one-click small-multiples feature, but a set of aligned charts fed from a dynamic, Table-based setup gets you there, and the payoff in readability over a six-line tangle is enormous. If you must keep one chart, gray out all series except the one or two the story is about — emphasis is a legitimate design tool; hoping the reader untangles six colors is not.

Sometimes the right chart is a table

Heresy on a charting site, but true: when your reader needs to look up exact values — a price list, a rate card, a handful of KPIs with targets — a well-formatted table beats any chart. Charts exist to show shape: trends, gaps, rankings, outliers. If there’s no shape to show, a chart is decoration around numbers that wanted to be read directly. The heat map is the productive middle ground — a table with intensity shading — which is why it earns its place in the matrix above.

Financial and stock data: a special case

Price-over-time data has its own conventions. A simple closing-price history is just a line chart, but open-high-low-close data has a purpose-built form — the candlestick chart, which Excel supports natively as a stock chart type. Use it when the within-period range is the point; use a plain line when only the trend matters. (Everything on this site about financial charts is educational charting technique, not investment guidance.)

The five wrong turns we see most often

Having reviewed a lot of business workbooks, the same handful of wrong-chart decisions account for most of the damage. Worth naming them explicitly.

Pie chart for time series. “Revenue share by month” as twelve pie slices. Months are a sequence, not a composition — this is line chart territory, full stop, and it’s the single most common miscast we encounter.

Column chart for 30+ categories. A column per customer for your whole customer list produces an unreadable comb. Either rank and cut (top 10 plus “all others” — honest and readable) or reach for a Pareto chart, which is purpose-built for exactly this “many contributors, few that matter” shape.

Stacked columns to compare a middle segment. Only the bottom segment of a stacked column sits on a common baseline; every segment above it floats, so comparing the middle series across columns is guesswork. If one segment’s trend is the story, give that segment its own chart — or use a line per segment.

Gauge farms. Six gauges across the top of a dashboard, each spending a quarter of the screen on one number. Covered in the KPI verdict above, but it earns a second mention because it’s the most space-expensive habit in dashboard design. A bullet chart stack shows the same six KPIs with targets and bands in a fraction of the area.

Scatter plot with time on the x-axis. If one axis is time, you want a line chart — scatters are for testing whether two measures relate. The reverse mistake (a line chart connecting unordered categories) is rarer but worse, since the connecting line implies a sequence that doesn’t exist.

None of these are exotic errors. They’re defaults chosen under deadline. The matrix at the top of this page exists so the right default is faster than the wrong one.

After you’ve chosen: five rules that make any chart better

Choosing correctly is most of the battle; these finish it.

Start bar and column baselines at zero, always. Length encoding only works from a common zero. Lines are exempt — position, not length, carries their meaning.

Sort by value unless order means something. Time stays chronological; survey scales stay in scale order; everything else sorts largest-first.

Label directly, kill the legend where you can. A label at the end of each line beats a legend the reader’s eye must shuttle to and from. On a two-series chart, a legend is an admission of defeat.

One message per chart. If your title needs “and,” you need two charts. The best chart titles state the finding — “Support costs fell 18% after the migration” — not the contents (“Costs by month”).

Delete non-data ink. Gridlines lighter or gone, borders gone, backgrounds white, no shadows. Every pixel that isn’t information competes with the pixels that are. If you produce a house style you like, save it as a .crtx chart template so you never rebuild it.

3D anything. 3D columns, 3D pies, 3D lines: the perspective distortion systematically misrepresents values (rear items look smaller, front pie slices look bigger), and it adds zero information. There is no dataset for which 3D is the right answer. None.

Pies beyond four slices. Covered above; worth repeating. Slice five onward is unreadable and should have been a bar.

Dual-axis charts with matching units. Two dollar series on two different dollar axes is a lie with extra steps. Same units, same axis — always.

Stacked areas with more than three series. Middle series ride on the wiggling series below them, so nobody can read their actual values. Use lines or small multiples.

Truncated-axis drama columns. Starting a column chart’s y-axis at 94 to make a 2% difference look like a cliff is the most common honesty failure in business charting. Column and bar baselines start at zero; if the meaningful variation is small, that’s what line charts (which don’t carry the zero-baseline convention) are for. Our axis scale guide covers taking manual control the honest way.

Radar charts for business KPIs. The enclosed area — the thing your eye reads — is an artifact of the arbitrary order you placed the axes in. Reorder the spokes, get a different-looking result from identical data. A bullet-chart stack or simple table beats it every time.

What about letting AI choose?

Fair question in 2026, since Copilot in Excel will happily suggest a chart type when you ask it to “visualize this.” Our take, expanded in the complete Copilot charts guide: Copilot has become a genuinely fast chart builder, but as a chart chooser it defaults to the safe and generic — clustered columns for nearly everything — and it will never push back on a misleading request. Chart choice is a three-second decision once you know the patterns on this page. Make the call yourself, then tell the AI exactly what to build. That order of operations gets you the speed of AI with none of its judgment gaps.

The two-minute method, summarized

Ask what your data is doing — comparing, trending, composing, building to a total, distributing, flowing, or relating. Take the matrix’s verdict. Sort your categories by value, label directly instead of relying on legends where you can, start bar and column baselines at zero, and delete anything on the chart that isn’t information. Then, if you want the build done in thirty seconds, hand the finished decision to Copilot or follow the step-by-step tutorial linked in the matrix.

Wrong chart, well made, is still wrong. Right chart, plainly made, wins every time.