Data Visualization

Choosing the wrong chart doesn't just make the visual harder to read. It often produces a conclusion the data doesn't actually support. A pie chart with 11 slices makes comparison impossible. A bar chart starting at $980K instead of $0 makes a 2% change look like a doubling. A line chart connecting monthly survey results implies trend continuity where none exists. The chart type is an argument — it claims a relationship between variables — and the wrong argument misleads even when the numbers are correct.

This guide is a working vocabulary for chart selection. For each major chart type, you'll find what question it answers, when it's the right choice, and the failure modes that come with it. The goal isn't to make you a data visualization designer. It's to make you fluent enough to read a dashboard critically, ask the right question about why a chart was made this way, and specify what you actually need when someone asks "how should I show this?"

We'll move in eight sections: comparison charts, trends over time, distributions, relationships, part-to-whole compositions, the specialized chart types that show up frequently in business but rarely in textbooks, dashboard design principles, and the chart pitfalls that mislead.

§1 Comparison charts — seeing differences between groups Foundational

Comparison charts answer "how do these categories differ?" They're the most common chart type in business because most business questions are comparisons: which region performed best, how does this quarter compare to last, which product line has the highest margin.

Bar charts (horizontal) and column charts (vertical) are the same chart, rotated. The convention that's evolved: use horizontal bars when your category labels are long (product names, department names, countries) or when there are many categories — the text is legible left-to-right. Use vertical columns for time-based categories (Q1/Q2/Q3/Q4, months, years) where the left-to-right time axis is natural, or for comparing fewer categories where vertical feels less cluttered.

The rule for both: start the axis at zero. A bar chart truncated at $980K to show three bars at $985K, $992K, and $997K makes a trivial 1% difference look enormous. The length of the bar is the argument — cutting the axis corrupts the visual claim. (Line charts and scatter plots don't have this rule; they represent position, not length.)

Grouped bar charts compare multiple series side-by-side within each category. Two or three groups work well; more than four and the visual becomes cluttered. When the number of series grows, consider a small multiples layout instead — separate charts per category, same scale — which allows side-by-side reading without crowding.

Lollipop charts are bar charts where the bar is replaced with a line and a dot. They carry the same information with less visual weight, which is useful when showing many categories where the bars would create a visually dense block. The tradeoff: they're less immediately recognizable than bars and require slightly more reading.

Bullet charts show a single metric against a reference (target, benchmark, prior period) on a single axis. They're ideal for dashboard KPI rows where you need to convey "current value + target + performance band" in minimal space. They replace the space-inefficient gauge charts that show up on many dashboards and convey exactly the same information.

Related glossary: dashboard, KPI.

§2 Trends over time — showing change Foundational

Time-series charts answer "how did this change?" They require time on the x-axis — and unlike comparison charts, the axis doesn't need to start at zero (because you're showing position, not length).

Line charts are the standard for continuous data over time — revenue per month, active users per week, stock price. The line implies continuity: a value at every point between the labeled points. This is usually true for rates and measurements taken continuously. It is not true for discrete survey scores, annual counts of discrete events, or anything where the values between labeled points are genuinely unknown. Using a line chart for three annual observations implies a smooth trajectory between them — which may not exist.

Area charts are line charts with the space below filled in. They work well for showing magnitude as well as direction — you can see both the trend and the volume. Stacked area charts show multiple series summing to a total. The failure mode: stacked areas make it hard to read individual series other than the bottom one. If you need to compare how individual series change, separate lines are cleaner.

Slope charts show how values change between exactly two time points for multiple series. They're compact and effective for comparisons like "how did each region's market share change between 2021 and 2025?" The two endpoints and the connecting slope convey direction and magnitude for each series simultaneously. They become cluttered beyond about 8-10 series.

The time-on-x convention is strong but has one important exception: when you want to show the relationship between two time-varying quantities, a scatter plot (§4) often works better than trying to layer both on a time axis.

Related glossary: mean, cohort analysis.

§3 Distributions — understanding the spread of values Building

Distribution charts answer "how are values spread?" They reveal information that averages and totals hide: whether data is symmetric or skewed, whether there are outliers, whether there are two clusters masquerading as one average.

Histograms divide a continuous range into equal-width bins and show how many values fall in each. They're the right chart for understanding the shape of a single continuous variable — revenue per customer, page load times, deal sizes, employee tenure. The y-axis is count (or frequency). The choice of bin width matters: too few bins and you lose the shape; too many and you see noise as structure.

Box plots (also called box-and-whisker plots) summarize a distribution's key statistics in a compact format: the central box spans the interquartile range (25th to 75th percentile), with a line at the median (50th percentile). Whiskers extend to the data range, and outliers are plotted as individual dots. Box plots shine when comparing distributions across multiple groups — four box plots side by side immediately show whether groups have similar medians, different spreads, or different outlier patterns. They're less intuitive for general audiences than histograms but more space-efficient when comparing groups.

Density plots (kernel density estimates) show the smoothed shape of a distribution as a continuous curve. They're similar to histograms but without the bin-width choice problem — the smoothing does that work. Overlapping density curves for two groups make a comparison of their shapes easy to read, which is where density plots have an advantage over two separate histograms.

Violin plots combine a box plot with a density curve, showing the full distribution shape plus the summary statistics in one visual. They require slightly more statistical literacy to read but are increasingly common in data team communications.

The key question when you see any distribution chart: is the distribution approximately symmetric (a classic bell curve shape), or is it skewed? Skewed distributions (most revenue distributions, most engagement distributions, most wealth distributions) mean the mean is pulled away from the typical case — the median is the more informative central tendency, and any analysis that assumes normality may be wrong.

Related glossary: mean, median, standard deviation, percentile, outlier, normal distribution.

§4 Relationship charts — seeing how variables connect Building

Relationship charts answer "how do these variables relate to each other?" They require two or more numeric variables measured at the same unit of observation (each dot represents one customer, one store, one campaign, etc.).

Scatter plots are the foundational relationship chart. Plot one variable on the x-axis, one on the y-axis, and each observation becomes a point. The pattern of dots shows the relationship: a cloud sloping up-right indicates positive correlation; sloping down-right indicates negative correlation; a random cloud indicates no linear relationship. Scatter plots make correlation visible and also make it easy to spot outliers — points that don't follow the pattern.

The failure mode of scatter plots: overplotting. When you have tens of thousands of observations, the dots stack on top of each other and the visualization is an opaque blob. Remedies: reduce dot opacity, use a 2D density plot, or bin the data before plotting.

Bubble charts extend the scatter plot with a third variable encoded as dot size. They're visually engaging but hard to read precisely — humans are poor at comparing areas. Use bubble size only for a variable that doesn't need precise comparison (market share category, rough magnitude tier). If you need the third variable to be precisely readable, encode it as color or facet the chart.

Heatmaps show the relationship between two categorical (or binned numeric) variables, with the value encoded as color intensity in each cell. They work well for calendar data (day-of-week × hour-of-day traffic patterns), for correlation matrices (how correlated are all pairs of variables in a dataset), and for cohort retention tables (month of acquisition × months since acquisition). Color scale choice is non-trivial: sequential scales (light to dark) work for data ranging from low to high; diverging scales (blue to white to red) work for data with a meaningful midpoint (correlation coefficients, net sentiment scores).

Related glossary: correlation, Pearson correlation, cohort analysis, A/B test.

§5 Part-to-whole charts — composition and proportion Building

Part-to-whole charts answer "what's the breakdown?" They show how a total divides into components. They require that the components are exhaustive (they add up to the whole) and mutually exclusive (each data point belongs to one component).

Pie and donut charts are the most familiar — and the most misused. The area of each slice represents its share of the whole. Humans are poor at comparing angles and areas, which means pie charts work only when: there are few categories (2-4 maximum), one category is clearly dominant, and exact proportions are less important than a gestalt sense of "roughly half" or "roughly a quarter." When you need readers to compare slice sizes — especially non-adjacent slices — a bar chart of percentages is almost always more readable. The donut variant (with a cutout center) is cosmetically different but functionally identical.

Stacked bar charts show part-to-whole composition across multiple groups or time points. Each bar represents a total; segments show components. They work well when there's a dominant component you want to track (the bottom segment is easy to read across bars; segments above it are harder), and when the number of components is small (3-5). Beyond that, individual segments become too narrow to label.

100% stacked bars normalize each bar to 100% and show the proportion shift across groups or time — useful for showing how the composition changes even when the total changes. The failure mode: you lose the absolute magnitude entirely. Revenue going from $1M to $10M with the same proportions looks identical to revenue staying flat in a 100% stacked bar.

Waterfall charts show how a starting value changes through additions and subtractions to reach an ending value. They're the standard chart for financial reconciliations: "We started the quarter at $12.4M ARR. New bookings added $2.1M. Expansions added $0.8M. Churn removed $1.3M. Net revenue retention added $0.4M. We ended at $14.4M." Each step floats — additions go up from the prior bar, subtractions go down. The final bar anchors to zero.

Treemaps divide a rectangle into nested tiles proportional to each component's size. They're compact and effective for showing hierarchy (categories within categories) and for giving a quick visual sense of where the "mass" is in a large dataset. They're poor for precise comparison — adjacent tiles of slightly different sizes are hard to distinguish.

Related glossary: ARR, churn, NRR.

§6 Specialized business charts — the formats that show up in decks Strategic

The charts in this section don't fit neatly into the prior categories but appear frequently in business contexts. Understanding their specific purpose — and their specific failure modes — prevents misuse.

Funnel charts show conversion through sequential stages, with each stage narrowing to represent drop-off. They're the standard chart for marketing conversion pipelines (impression → click → lead → opportunity → close), onboarding flows, and multi-step processes. The stages are ordered, and the area or length of each stage represents the count or percentage remaining.

The failure mode: funnel charts imply every entry at stage 1 progressed through the stages in sequence. If your pipeline has re-entry (a churned customer who re-engages) or skip-ahead (a direct response that enters at stage 3), a funnel chart distorts the actual flow. Sankey diagrams (below) handle branching and re-entry better.

Gantt charts show project schedules, with tasks on the y-axis and time on the x-axis. Each task is a horizontal bar spanning its start and end date. Dependencies are shown with arrows. They're the standard tool for project planning and stakeholder communication about timelines. The failure mode: they look more precise than they are. A task bar from May 1 to June 15 implies the work is linear and bounded — which is rarely true for knowledge work. Gantt charts are better for communicating sequencing and handoffs than for tracking actual completion against the plan.

Sankey diagrams show flow between categories, with the width of the flow proportional to its volume. They're excellent for showing how traffic, revenue, or users move between states: where web sessions originate and what pages they flow through; how customers move between pricing tiers; how budget allocates from source to destination. They're more complex to read than most charts and should be reserved for situations where the flow pattern itself is the key insight — not for audiences who aren't already comfortable with charts.

Cohort retention heatmaps (covered briefly in §4) deserve expansion here because they're so widely used in SaaS, subscription, and consumer product contexts. A retention heatmap is a grid: rows are cohorts (by acquisition month or quarter), columns are "time since acquisition" periods (month 1, month 2, ...). Each cell shows the retention rate for that cohort at that time step. Reading across a row: one cohort's retention curve over time. Reading down a column: how retention at a given age (month 2, say) is trending across cohorts — are newer cohorts retaining better or worse at month 2 than older ones? The diagonal (all cells at the same calendar month) shows what happened in a given month to all currently-active customers simultaneously — a signal of seasonal or marketing effects.

Related glossary: conversion rate, funnel, cohort analysis.

The right chart is rarely obvious on the first try, and that's fine. What matters is starting with the question — what comparison, trend, distribution, relationship, or composition am I trying to show? — before choosing the chart type. The chart serves the question, not the other way around.

The charts that cause the most damage in business communication are usually well-intentioned but made with the wrong framing: "this is the data I have, what chart can I make?" rather than "this is the decision I'm trying to support, what information does the reader need?" Starting from the decision almost always produces a cleaner, more honest chart — and it's the question your colleagues, managers, and board members are asking anyway.

See also: Statistical Concepts Guide — the underlying statistical methods for distributions, correlations, and significance that determine which chart type is appropriate; Data Foundations Guide — the data infrastructure that produces the data being visualized.

§7 Dashboard design principles — building views people actually use Strategic

A dashboard is not a collection of charts. It's an answer to a question. That distinction sounds minor until you look at how most dashboards fail: a grid of twelve charts, no visible hierarchy, no stated question, and an audience that loads it once, decides there's nothing actionable, and never returns. The question-first principle is the first thing to establish before building anything — who is the audience, and what specific question does this dashboard exist to answer? "Marketing performance" is not a question. "Is our paid acquisition CAC moving toward or away from target this quarter?" is a question.

Information hierarchy is the visual expression of priority. The most important number on the dashboard should be the physically largest element — a KPI card with the current value, the trend direction, and comparison to target. Supporting context goes below. Exploratory detail goes last. This mirrors how readers actually navigate a screen: they scan for the headline before they read the article. A dashboard that buries the most important number in row three of a table has a hierarchy problem, not a data problem.

Operational dashboards and analytical dashboards are different products and should be designed differently. Operational dashboards are for a narrow, specific audience (an ops manager, an on-call engineer, a support-team lead) who needs real-time or near-real-time status on a small number of metrics they action directly. The design imperative is speed to decision: large numbers, clear thresholds, minimal decoration. Analytical dashboards serve a broader audience exploring trends over time — they need more chart types, filtering capability, and contextual comparators (vs prior period, vs target, vs segment). Designing one and calling it both is how dashboards become unusable.

Progressive disclosure is the pattern for analytical dashboards: summary view (headline KPIs) → drill-down (trend and segment breakdown) → detail (transaction or record level). Every layer should be accessible, but the default view should answer the question without requiring the user to dig. Most users never click past layer one; the few who do are the analysts who need the detail. Build for the majority default, accommodate the minority path.

The 5-second test is a useful self-check before finalising a dashboard: show it to someone unfamiliar with the data for 5 seconds, then hide it and ask "what was this dashboard about?" and "what should the viewer do next?" If they can't answer either question, the hierarchy and question framing aren't working. The test is informal but tends to catch the most expensive design mistakes before they ship.

Why most dashboards fail comes down to a small set of recurring errors: building before defining the question; including every available metric instead of the useful ones; choosing chart types that require explanation; and designing for the builder's mental model of the data rather than the viewer's actual workflow. The most over-built dashboards are almost always built by the people who know the data best — expertise produces the temptation to show everything.

Related glossary: dashboard, KPI, data visualisation, operational analytics, business intelligence, drill-down.

§8 Chart pitfalls — common mistakes that mislead Strategic

Every chart that misrepresents data has a mechanism. Understanding the mechanism is more useful than memorising a list of forbidden chart types — because the same mechanism that makes a truncated Y-axis misleading on a bar chart also makes a zoomed-in line chart misleading on a dashboard. The underlying error is the same: the visual encoding implies a magnitude that the data doesn't support.

Truncated Y-axes are the most common quantitative distortion. When a bar chart's Y-axis starts at 95 instead of 0, a 2% change in a metric looks like a 400% change visually. This is almost always done to "make the trend more visible," but it conflates signal amplification with misrepresentation. The fix: start Y-axes at zero for magnitude comparisons (bar charts, area charts). The exception: if you're explicitly charting change over a narrow range and labelling it clearly as such, truncation is defensible — but it should be labelled, not assumed.

Dual Y-axes are seductive because they let you show two metrics on one chart — revenue and headcount on the same time series, for example. The problem is that the visual impression of correlation depends entirely on how the two axes are scaled, which is an arbitrary choice by the chart builder. By adjusting the scale of either axis, you can make any two time-series look perfectly correlated or completely uncorrelated. The fix: use separate charts when two metrics need different scales.

Pie charts with too many slices fail because human perception of angles is poor. We can reliably distinguish two slices; past five, the comparison requires squinting at the legend. Any pie with more than four or five segments should become a sorted bar chart, which uses length — the most accurate visual encoding for quantity — instead of angle. A related failure: using pie charts to compare across multiple groups (a grid of pies). This is almost impossible to read accurately; use a stacked bar chart instead.

3D charts add a third visual dimension to two-dimensional data. The distortion this introduces is not cosmetic — perspective foreshortening systematically makes near elements look larger than far elements, skewing the apparent relative magnitudes of whatever is being shown. There is no case in business data visualisation where a 3D chart is more accurate than a 2D chart. The fix is always to remove the third dimension.

Overloaded line charts with more than five lines produce a spaghetti effect where individual series become untrackable. The fix is usually to show one or two primary lines and small multiples for the rest — one mini chart per series, same scale, arranged in a grid. Comparison across small multiples is easier than untangling a spaghetti chart.

Cherry-picked date ranges are a form of selective truth-telling. Showing a three-month chart that captures a positive trend while omitting the prior six months of decline is technically accurate data, visually misleading framing. The fix: show enough history to provide context, and label clearly when the selected range was chosen for a specific reason.

Each pitfall distorts the same thing: the viewer's sense of magnitude or trend direction. The mechanism is visual — the data itself may be accurate. That's what makes these hard to spot.

Related glossary: data visualisation, chart type, Y-axis, correlation vs causation, misleading statistics, small multiples.