Chapter 9: Visuals and Data: Figures, Tables, Charts, and How to Write About Them

"Graphical excellence is that which gives to the viewer the greatest number of ideas in the shortest time with the least ink in the smallest space." — Edward R. Tufte, The Visual Display of Quantitative Information (1983)

Chapter Overview

A scientist drops a chart into a report and writes underneath it: "Figure 3 shows the results." They believe they have communicated. They have not. They have parked a picture and a caption next to each other and trusted the reader to do the work of figuring out why either one is on the page. The reader, scanning, sees a tangle of bars and a caption that says nothing, and moves on—taking with them exactly none of the insight the scientist spent three weeks earning. The data was right. The chart was accurate. The communication failed at the last inch.

This chapter is about that last inch. A figure, a table, and a sentence of prose are three different tools for carrying information into a reader's head, and most technical writers use all three badly—not because they can't make a chart (software makes charts) but because they treat the visual as the message instead of treating the interpretation as the message. The single most under-taught skill in technical communication is the interpretive caption: a caption that tells the reader what the figure means, not merely what it is. We will spend much of this chapter on it, because once you can write one, your data starts to land. This builds directly on Chapter 4's lesson that the reader scans (so a figure must carry its own point for the scanner who reads only captions) and Chapter 3's lesson that every element must earn its place (a figure that duplicates the prose earns nothing). By the end of this chapter, you will be able to decide whether information belongs in a figure, a table, or a sentence; design a chart that shows the data and not the decoration; write a caption that interprets; and weave a number into prose so it leads with its meaning.

The stakes are not always academic. On a cold night in January 1986, engineers tried to use charts to stop a rocket launch, and the charts—accurate, detailed, sincere—did not make the one fatal relationship visible. We met that failure in Chapter 4 as a structural problem; here we meet it again as a data-display problem, through Edward Tufte's analysis of the actual charts. The lesson converges: a visual that does not make its point unmissable is a visual that has failed, and when the stakes are high enough, "failed to make the point" and "failed to prevent a disaster" are the same sentence. We keep strictly to the verifiable historical record.

In this chapter, you will learn to:

• Choose deliberately among a figure, a table, and prose, based on what the reader needs to do with the information.
• Strip a chart to its data using Tufte's principles—maximize data-ink, delete chartjunk—so the signal is not buried in decoration.
• Write captions that interpret ("Onboarding, not price, drives churn") rather than label ("Churn by cohort").
• Write about data in prose so the finding leads and the figure supports ("Revenue fell 23% in Q4 (Figure 3), driven by enterprise non-renewals"), never the limp "Figure 3 shows revenue."
• Spot and fix the common failures: redundant figures, under-captioned tables, misleading axes, and findings the reader has to excavate.

📊 All three tracks need this chapter. 📕 Engineering/Science: your figures and tables are the load-bearing evidence of every report and paper; §9.3–§9.5 are core, and Chapter 18 (posters) and Chapter 13 (results sections) build on them. 📗 Software/CS: dashboards, data memos, and architecture docs live or die on captions and labels—§9.6 feeds Chapter 27 (data memos) and Chapter 32 (diagrams). 📘 Business/Professional: the chart in your deck or one-pager persuades only if its title states the takeaway—§9.5 (interpretive titles) is the highest-leverage section for you, and it returns in Chapter 30 (slides).

9.1 Three Tools, Three Jobs: Figure, Table, or Prose?

Before you make a chart, ask a question almost nobody asks: does this information even want to be a chart? You have three tools for delivering quantitative or structured information—a figure (a chart, graph, plot, diagram—anything visual showing relationships in data), a table (rows and columns of exact values), and prose (sentences). Each is best at a different job, and the most common data-display mistake is reaching for the wrong one out of habit. Charts feel impressive, so people chart things that should be a sentence; tables feel rigorous, so people tabulate things the reader will never look up.

Start with the principle, then we'll make it concrete. A figure shows a pattern, a relationship, a trend, a comparison, a shape. Use a figure when the shape of the data is the point—when it rises, falls, clusters, correlates, or diverges, and you want the reader to see that shape in one glance. The eye is extraordinary at perceiving slopes, peaks, and outliers; a figure offloads work from the reader's slow verbal reasoning onto their fast visual perception. A table gives exact values the reader needs to look up or compare precisely. Use a table when the specific numbers matter—when a reader will want to find the value for a particular row, compare exact figures, or check a number you'll be held to. Prose carries a single number, a simple comparison, or the reasoning that connects data points. Use a sentence when there are only one or two numbers, or when the relationship between numbers is an argument (which, as Chapter 4 taught, prose carries and bullets and tables cannot).

Watch the same information delivered three ways, and notice that only one of them fits.

❌ A sentence forced into a chart. A quarterly report contains a full bar chart titled "Total Customers," with a single bar of height 4,812, axis gridlines, a legend, and a caption: "Figure 2. Total customers in cohort."

✅ The same information as prose: "The analysis covers all 4,812 new accounts opened in the last 90 days."

Why it's better: One number is not a pattern; there is nothing to see. A bar chart of a single value wastes a quarter-page, makes the reader decode axes to retrieve a figure a sentence states instantly, and signals that the writer reached for "chart" by reflex. When you have one number, write the number.

Now the opposite error:

❌ A pattern buried in a table. A report presents a 12-row table of monthly churn rates (Jan 2.1%, Feb 2.0%, … Dec 4.0%) and writes: "Table 4 shows monthly churn." The reader is left to mentally plot twelve numbers to notice they are climbing.

✅ The same data as a figure: a line chart of churn over twelve months, the line sloping clearly upward, captioned "Churn nearly doubled over the year, from 2.1% in January to 4.0% in December (Figure 4)."

Why it's better: The point is the trend—churn is rising—and a trend is a shape. A table makes the reader do the plotting in their head; a line chart shows the rise instantly. If the reader will never need the exact February value but absolutely needs to grasp "it's getting worse," that's a figure, not a table.

So when would you use the table? When the exact monthly values are themselves load-bearing—when finance needs to reconcile each month, or a later section refers to "the 3.4% spike in September." Then give both: a figure for the shape in the body, a table for the values in an appendix. The deciding question is never "which looks more professional?" It is "what will the reader do with this—see a shape, look up a value, or follow an argument?" Match the tool to the task.

🔄 Check Your Understanding You need to report that your new caching layer cut average page-load time from 1,400 ms to 180 ms. Figure, table, or prose—and why?

Answer

Prose. It's two numbers and a simple before/after comparison, and the relationship (the cache caused an 87% drop) is an argument prose carries cleanly: "The caching layer cut average page-load time from 1,400 ms to 180 ms, an 87% reduction." A two-bar chart would be decoration around a sentence; a table would be a two-cell table no one looks up. Reserve the figure for when you have a shape (load time across many endpoints, or over time) and the table for when readers must retrieve exact per-endpoint values.

A practical default to carry out of this section: prose for one or two numbers; figure for a shape; table for exact lookup—and never two of them for the same fact (that redundancy is its own mistake, §9.7). When in doubt, ask what the reader's eyes are for in that moment, and give them the tool that does the work.

9.2 What a Caption Is For: The Threshold Concept

Here is the idea that changes how you handle every figure for the rest of your career, so we'll mark it as a threshold and slow down.

🚪 Threshold Concept: A figure does not speak for itself.

How most writers think about figures: the chart contains the information. The reader will look at it, perceive the data, and draw the right conclusion. The caption's job is to label the figure so the reader knows what they're looking at—"Figure 3: Churn by cohort." The insight lives in the picture; the words just name it.

How figures actually work: a chart shows data, but data is not a conclusion. A reader—especially a scanning reader, especially a non-expert—looks at your bars and does not automatically extract the point you extracted after weeks of analysis. They see shapes and need to be told what the shapes mean. The caption is not a label; it is the figure's argument. The most important sentence attached to any figure is the one that says what it means. When you cross this threshold, you stop writing captions that name figures and start writing captions that interpret them—and your data starts to persuade.

Why does this matter so much? Because of two facts you already know. First, the reader scans (Chapter 4): many readers will look only at your figures and their captions, never reading the surrounding prose. For those readers, the caption is the entire message. If it merely labels, they get a labeled picture and no insight. Second, the curse of knowledge (Chapter 2): you, who did the analysis, look at the chart and see the conclusion because you already know it. The reader does not have your context, and the shape that screams "onboarding is the problem!" to you is, to them, just some bars of different heights. The caption is where you hand the reader the conclusion they cannot reliably reach alone.

This single shift—caption as interpretation, not label—is the spine of the whole chapter. We'll spend §9.5 learning to write interpretive captions in detail. For now, install the instinct: whenever you place a figure, ask not "what is this a picture of?" but "what do I want the reader to conclude from it?"—and then make the caption say that.

🔍 Why Does This Work? Why does an interpretive caption help even a reader who could have figured out the figure on their own? Think before reading on.

Because of cognitive load and time. A reader who can derive your conclusion from the chart still has to spend effort doing it—holding the bars in mind, comparing them, reasoning to a takeaway—and a scanning reader will not spend that effort; they'll glance and move on. An interpretive caption does the reasoning for them, so the insight transfers in one read instead of requiring a small act of analysis. It also guarantees they reach the conclusion you intended rather than a different one they might infer. You're not insulting the reader's intelligence; you're respecting their time and removing the chance they take away the wrong message—or none. (This is the same logic as conclusion-first structure in Chapter 4: give the reader the takeaway, then let the evidence support it.)

9.3 Designing the Figure: Tufte, Data-Ink, and Chartjunk

A caption can only interpret a figure the reader can actually read. So before captions, the figure itself has to be clean. The foundational thinker here is Edward Tufte, whose 1983 book The Visual Display of Quantitative Information (Tier 1) gave the field two ideas you should carry everywhere: data-ink and chartjunk.

Data-ink is the ink (or pixels) that represents actual data—the line in a line chart, the tops of the bars, the points in a scatterplot. Non-data-ink is everything else—gridlines, backgrounds, borders, 3-D effects, redundant labels, decorative shading. Tufte's principle, the data-ink ratio, says: of the total ink in a graphic, maximize the share that is data-ink. Erase everything that isn't carrying information. Every gridline, box, and gradient that doesn't help the reader see the data is competing with the data for attention—and losing the reader's attention is the whole cost.

Chartjunk is Tufte's term for the visual noise that adds no information: 3-D bars that distort the very heights they're supposed to show, heavy gridlines, busy background images, gratuitous color, clip-art, "moiré" patterns of dense hatching. Chartjunk doesn't just waste ink; it actively obscures the data and can mislead. A 3-D pie chart tilts the slices so the front ones look bigger; a textured fill makes a bar's height ambiguous. The decoration that was meant to make the chart "look nicer" makes it lie.

Let's strip a junky chart down, in words (since this is Markdown, I describe each figure fully, with alt-text, per the book's accessibility standard).

❌ Before — a chart drowning in non-data-ink. Figure 9.1a (described): A bar chart of churn rate for three onboarding-time cohorts. The bars are rendered in 3-D with drop shadows, sitting on a blue gradient background. Heavy dark gridlines cross the plot every 2%. Each bar is a different bright color, explained by a legend off to the right (which merely repeats the x-axis labels). The bars are filled with diagonal hatching. The title reads "CHURN ANALYSIS." Because of the 3-D tilt, it's hard to read each bar's true height against the axis. Alt-text: a cluttered 3-D bar chart whose decoration—shadows, gradient, hatching, redundant legend, heavy gridlines—makes the actual bar heights hard to read.

✅ After — the same data, decoration removed. Figure 9.1b (described): A flat 2-D bar chart, three bars, no background, no border, no legend. Light gray axis line only; faint horizontal reference at 0% and at the top value. Bars are a single muted color, directly labeled with their values above them: "<7 days: 3%", "7–21 days: 9%", ">21 days: 22%." The title states the finding (see §9.5): "Slow onboarding, not price, drives churn: customers who take 3+ weeks to first value cancel 7× as often." Alt-text: a clean bar chart showing churn rising with onboarding time—3% for under 7 days, 9% for 7–21 days, 22% for over 21 days—with each bar directly labeled and a title stating the takeaway.

Why it's better: Every change in 9.1b removed something that wasn't data and added clarity. Killing the 3-D tilt lets the reader read true heights. Deleting the redundant legend removes a thing the reader had to cross-reference for no information. Direct value labels mean the reader never has to trace a bar to a gridline. The single color stops the rainbow from implying the three cohorts are unrelated categories rather than points on one scale. The result carries more information with less ink—Tufte's definition of graphical excellence. The "before" looked busy and "designed"; the "after" looks plain and communicates.

A few practical rules that fall out of data-ink and chartjunk, usable today:

• Delete gridlines or make them faint. They're scaffolding; the reader rarely needs to read exact values off them, and if they do, label the points directly.
• Avoid 3-D for 2-D data. Almost never use a 3-D bar, pie, or area chart. The third dimension carries no information and distorts the two that do.
• Label directly, drop legends when you can. A legend forces the eye to bounce between the data and a key. If you can put the label on the line or bar, do.
• Pick the right chart for the relationship. Trends over time → line. Comparison across categories → bar. Correlation between two variables → scatter. Parts of a whole → usually a bar or a sentence, rarely a pie (humans compare angles poorly; a pie with more than a few slices is unreadable).
• Start bars at zero. A bar's length encodes its value, so a truncated axis lies about proportions (more in §9.7). Line charts can sometimes start elsewhere to show a trend, but say so.

🧩 Productive Struggle Before reading the answer, redesign this in your head. A team wants to show that one of their five servers handles far more traffic than the others. They've made a pie chart with five slices in five colors, a legend, percentages on each slice, and a 3-D tilt. Name two things wrong with this choice and what you'd make instead.

Answer

Problems: (1) A pie chart is a poor tool for comparing five magnitudes—humans judge angles and areas badly, so "far more" is hard to see; a bar chart, where the eye compares lengths along a common baseline, shows the disparity instantly. (2) The 3-D tilt distorts the slices (front slices look larger), so the very comparison the chart exists to make is corrupted. The legend also forces cross-referencing that direct labels would avoid. Fix: a flat horizontal bar chart, servers sorted largest-to-smallest, each bar directly labeled with its load, so the dominant server is obviously the longest bar. Better still, a caption that states it: "Server 3 handles 61% of traffic—more than the other four combined." The shape and the words both carry the point.

[📍 Good stopping point — you now have the figure-vs-table-vs-prose decision and a clean-figure discipline. The rest of the chapter is the writing: captions that interpret, and prose that carries data.]

9.4 Exploratory vs. Explanatory: The Chart You Make for Yourself Is Not the Chart You Publish

A distinction that prevents a great deal of bad data communication: the graphics you make to understand your data are not the graphics you make to explain it to a reader. Call them exploratory and explanatory graphics.

Exploratory charts are for you, the analyst, during analysis. You make dozens of them, fast and ugly, to find what's going on—every variable against every other, histograms, raw scatterplots with all 4,812 points, default colors, no labels, no captions. Their audience is one expert (you) who already holds all the context. Speed and breadth matter; polish doesn't.

Explanatory charts are for the reader, in the final document. There should be few of them, each chosen because it carries one specific point the reader needs, each cleaned (§9.3) and captioned to interpret (§9.5). Their audience is someone who was not in your head during the analysis.

The classic failure is publishing an exploratory chart. You found the insight in a dense, multi-series, default-styled plot, and because you can read it (you made it), you paste it straight into the report. To the reader it's an unreadable thicket—twelve overlapping lines, a legend of cryptic variable names, no indication of which line matters. You're showing the reader your search instead of your finding. This is the curse of knowledge wearing a lab coat: the chart that's legible to you because you remember building it is illegible to the reader who wasn't there.

❌ Before (exploratory chart, published): A scatterplot of all 4,812 accounts, churn (0/1) on the y-axis jittered into two clouds, time-to-first-value on the x-axis, default blue dots, no trend line, caption "Figure 5: TTFV vs churn." The relationship is in there, but a non-analyst sees two smears of dots and learns nothing.

✅ After (explanatory chart, built for the reader): The same relationship reduced to three labeled bars (the cohort chart from 9.1b): churn for <7 days, 7–21 days, >21 days. The 4,812 raw points become three numbers the reader can grasp, with an interpretive caption. The exploratory scatter, if needed at all, moves to an appendix for the one reviewer who wants the raw shape.

Why it's better: The explanatory version answers "what should the reader conclude?"—churn climbs steeply with onboarding time—rather than "what does my raw data look like?" It buckets the cloud into a comparison the eye reads at a glance and the caption can interpret. The exploratory scatter was correct and useless to the audience; the bar chart is a deliberate act of communication.

The habit to build: when a chart is ready to go into a document, ask "did I make this to find something, or to show something?" If you made it to find something, it is almost certainly the wrong chart for the reader—rebuild it as the simplest figure that carries the one point, then write the caption.

🔄 Check Your Understanding Your notebook has a correlation heatmap of 20 variables that helped you spot that two features predict churn. Should it go in the report to leadership? Why or why not?

Answer

No—it's exploratory. A 20×20 heatmap helped you find the signal, but leadership doesn't need (and can't read) the full correlation matrix; they need the finding it led you to. Publish an explanatory figure of just the two features that matter, with an interpretive caption ("Two factors predict churn: onboarding time and support-ticket count"). Keep the heatmap for the methods appendix or the technical peer review (Priya), where a reader with context might want it. Match the graphic to the audience and the purpose, exactly as you match the document (Ch 2).

9.5 Captions That Interpret: The Most Under-Taught Skill

Now the heart of the chapter. We established (§9.2) that a caption should interpret, not label. Here is how to actually write one—and we'll use Dana Whitfield's churn work as the running example, because data captions are exactly the place her analysis succeeds or fails with the reader.

Recall Dana, the data scientist from Chapter 2: she's analyzed customer churn for Renée Okafor, the VP of Marketing, and found that new customers who take more than three weeks to get value from the product cancel about seven times as often as those who get there within a week. The leak is onboarding, not price or product. Dana has a clean bar chart (9.1b). Now she has to caption it. Watch the caption climb from useless to excellent.

❌ Level 0 — the non-caption. "Figure 3."

A figure number and nothing else. The reader has no idea what it shows or why it's there. Distressingly common.

❌ Level 1 — the label (describes what it IS). "Figure 3. Churn rate by onboarding-time cohort."

This names the axes. It tells the reader what kind of picture they're looking at. It does not tell them what to conclude. A scanning reader reads this, looks at three bars, and has to do the interpretive work themselves—and a non-expert like Renée may not. This is where most technical writers stop, and it is not enough.

🟡 Level 2 — the observation (describes what it SHOWS). "Figure 3. Churn rises with onboarding time, from 3% for customers reaching first value within a week to 22% for those taking over three weeks."

Better—now the caption states the pattern, so the reader who only reads captions gets the trend. But it still stops at description. It says what happened, not what it means or what to do about it. We're close.

✅ Level 3 — the interpretation (states what it MEANS). "Figure 3. Onboarding speed, not price, drives churn: customers who reach first value within a week cancel at 3%, versus 22% for those who take over three weeks—a 7× difference that points to the first-week experience as the place to intervene."

Now the caption carries the conclusion. It names the cause ("onboarding speed, not price"), quantifies the effect (3% vs. 22%, 7×), and points to the action (intervene in the first week). A reader who reads only this caption and the chart leaves with Dana's actual finding and its implication. That is an interpretive caption.

Why Level 3 is the target: It does the reader's reasoning for them and guarantees the right takeaway. The figure shows the bars; the caption supplies the meaning the bars can't state. Notice it isn't long-winded—it's one sentence that happens to do three jobs (cause, magnitude, implication). The progression Level 1 → Level 3 is the move to internalize: from what it is → what it shows → what it means.

A reliable template for getting to Level 3, when you're stuck:

[The finding/claim], shown by [the key comparison or number]: [the specific values], which means [the implication for the reader].

Run Dana's data through it: "Onboarding speed drives churn [claim], shown by the gap between fast and slow cohorts [comparison]: 3% versus 22% [values], which means retention effort should move to the first week [implication]." You won't always use every slot, but the template forces the two things weak captions omit—the claim and the implication.

Three more rules for good captions, with quick examples:

1. The caption should be self-contained. A reader who jumps straight to the figure (scanners do this constantly) should understand it without hunting through the body text for context. Define the cohort, name the units, state the takeaway—right there in the caption.

❌ "Figure 4. Results after the change." (After what change? What results?) ✅ "Figure 4. After moving onboarding emails to day 1 (from day 7), week-one activation rose from 41% to 68%."

2. Put the number that matters in the caption, not just the chart. If the reader takes one number away, make sure it's in words, not only encoded in a bar they have to measure.

❌ "Figure 5. Quarterly revenue." ✅ "Figure 5. Revenue fell 23% in Q4—the first decline in nine quarters—driven by enterprise non-renewals."

3. State the so-what, not just the what. Borrow Chapter 3's "so what?" test and apply it to captions: if your caption could be followed by a reader thinking "okay… and?", it hasn't reached interpretation.

❌ "Figure 6. Error rate over time." (And?) ✅ "Figure 6. The error rate tripled the night of the deploy and stayed elevated for six hours—evidence the regression was introduced by the release, not the load spike."

✏️ Try This Take this label caption and rewrite it as an interpretive (Level 3) caption. You may invent a plausible takeaway. "Figure 7. Support tickets by category." Imagine the chart shows that "Login/Access" tickets are 3× any other category.

One possible rewrite

"Figure 7. Login and access problems dominate support load—three times the volume of any other category—making authentication the highest-leverage area to fix." It names the finding (login dominates), gives the magnitude (3×), and states the implication (fix auth first). A reader who sees only this caption knows where to send engineering effort.

This is the skill that Chapter 27 (data memos) and Chapter 18 (posters) and Chapter 30 (slides) all build on, so it's worth over-practicing now. The shortcut that makes captions excellent forever: write the caption you'd want if you were the busy reader who reads only the captions.

9.6 Writing About Data in Prose: Lead With the Finding, Not the Figure

Captions handle the reader who looks at the figure. But you also write about your data in the running text, and there's a parallel skill—and a parallel failure—there. The failure is pointing at the figure instead of stating the finding.

The canonical weak move:

❌ Before: "Figure 3 shows the relationship between onboarding time and churn rate."

This sentence is about the figure rather than about the data. It tells the reader a figure exists and that it shows something, then makes them go look and figure out what. It's the prose equivalent of a Level-1 label caption. Worse, it wastes the topic-sentence slot—the most valuable real estate in the paragraph (Chapter 4)—on an administrative announcement instead of an idea.

The fix is to lead with the finding and demote the figure reference to a parenthetical citation:

✅ After: "Customers who take more than three weeks to reach first value churn seven times as often as those who get there within a week (Figure 3)—the strongest signal in the data that onboarding, not price, drives cancellation."

Why it's better: The sentence now states the finding as its main clause; the figure becomes a citation (Figure 3) the way you'd cite a source—evidence for the claim, not the subject of the sentence. The reader gets the insight from the prose itself and can verify it in the figure if they wish. This is the difference between prose that carries the analysis and prose that merely gestures at a chart.

The pattern generalizes into a habit you can apply mechanically:

Weak: "[Figure/Table N] shows/illustrates/presents [topic]." Strong: "[The finding, as a complete claim with the key number] ([Figure/Table N])."

More transformations, across fields, so you see the shape:

❌ "Table 2 presents the benchmark results for the three databases." ✅ "Postgres outperformed both alternatives on write-heavy workloads, sustaining 14,000 inserts/sec versus 9,000 and 6,500 (Table 2)."

❌ "The graph in Figure 8 displays patient recovery times across the two treatment groups." ✅ "Patients in the treatment group recovered a median of four days faster than controls (Figure 8), a difference that held across age bands."

❌ "Figure 9 illustrates our monthly active users over the past year." ✅ "Monthly active users grew 40% over the year, with the steepest gains after the March redesign (Figure 9)."

Notice three things these strong versions share. They lead with the claim (finding first, Chapter 4's inverted pyramid at the sentence scale). They include the load-bearing number in words, not only in the figure. And they relegate the figure to a citation, usually in parentheses at the end. That trio—claim, number, citation—is the formula for writing about data well.

Two refinements. First, interpret, don't just report, the number. "Revenue fell 23%" reports; "Revenue fell 23%, the first decline in nine quarters, driven by enterprise non-renewals" interprets—it contextualizes (first in nine quarters) and explains (driven by). The reader needs to know not just that the number moved but whether it's surprising and why. Second, don't narrate the chart's geometry. Avoid "the line goes up and then plateaus and then dips slightly before rising again"; the reader can see the line. Say what the shape means: "Growth stalled mid-year before the redesign reignited it." Describe the meaning, not the wiggles.

🔄 Check Your Understanding Rewrite this sentence to lead with the finding: "Figure 4 shows the distribution of response times for our API." Assume the data shows most responses are fast but a long tail exceeds two seconds.

Answer

"Most API responses return under 200 ms, but a long tail exceeds two seconds (Figure 4)—the slow tail, not the median, is what users actually feel." The finding (fast median, slow tail) leads; the figure is demoted to a citation; and it interprets the so-what (the tail is what hurts users). Compare the original, which announced a figure and a topic and made the reader do the rest.

9.7 The Narrative of Data: Interpret, Don't Just Present

Step back from individual captions and sentences to the whole act of presenting data. The deepest mistake in technical writing about data is not a bad chart or a weak caption; it is treating the job as presentation when the job is interpretation. Presenting data means showing the reader your numbers and figures and trusting them to find the story. Interpreting data means telling the reader the story and using the numbers and figures as evidence.

The presenter writes: "Here is the revenue chart. Here is the churn table. Here is the cohort breakdown. Here is the regional split." Four exhibits, no thread. The reader is handed a pile of true facts and left to assemble meaning—which they won't, because that's the analyst's job, not the reader's. The interpreter writes: "We're losing new customers faster than we're keeping them, and the cause is onboarding. Here's the evidence." Then every figure that follows is in service of that claim, in an order that builds it. Same data; one is a warehouse, the other is an argument.

Data has a narrative, and your job is to tell it. A useful shape, borrowed from Chapter 4's structures, is: the finding (the headline) → the evidence (figures and numbers, in an order that builds the case) → the implication (so what should we do). Lead with the conclusion (BLUF), support it with well-captioned figures each carrying one point, and end on the action. The figures are not the message; they are the citations for a message you state in words.

❌ Before (presentation — a warehouse of exhibits): "This report contains our Q4 customer analysis. Figure 1 shows total customers. Figure 2 shows churn by month. Figure 3 shows churn by cohort. Table 1 shows revenue by segment. Figure 4 shows onboarding completion rates. Figure 5 shows support ticket volume. These figures show the state of our customer base in Q4."

✅ After (interpretation — an argument with evidence): "We have a growing churn problem, and it starts in the first week of the customer's life—not at renewal, and not over price. Churn rose from 2.1% to 4.0% over the year (Figure 2). The cause is visible in the cohorts: customers who reach first value within a week churn at 3%, while those who take over three weeks churn at 22%—a 7× gap (Figure 3). That slow group is large because only 41% of new customers complete onboarding (Figure 4). The recommendation follows directly: move retention investment upstream into the first-week experience, where the leak actually is."

Why it's better: The "before" is an inventory—it lists what exists and asserts nothing. Every "Figure N shows" is a Level-1 label in prose form. The reader finishes knowing that figures exist and not what they mean. The "after" makes a claim (churn is rising and onboarding is the cause) and marshals exactly the figures that support it, each cited as evidence, in an order that builds the case to a recommendation. Two of the original five figures (total customers, support tickets) don't appear—because they didn't serve the argument, and an interpreter cuts exhibits that don't earn their place (Chapter 3, theme 6). The "after" is shorter, clearer, and actually persuades.

This is theme 5 (structure serves the reader) and theme 3 (clarity isn't the enemy of precision) meeting over data: you don't dump every number you have; you select and order the evidence to make the true finding unmissable. Precision survives—every number in the "after" is real and checkable—but it's organized into meaning. That's the difference between a data analyst who gets ignored and one who changes a decision. We develop this into the full data-memo discipline in Chapter 27, where Dana's memo gets rebuilt through three drafts; here, internalize the core move: don't present your data, interpret it—state the finding, then use the figures as evidence.

🧩 Productive Struggle You have three figures: (A) a line chart showing server response time spiking at 8 p.m. nightly, (B) a bar chart showing the 8 p.m. spike correlates with a nightly batch job, (C) a table of all batch-job schedules. You need to tell engineering leadership that the nightly report job is degrading performance and should be rescheduled. In one or two sentences, write the opening of your interpretation (the finding), and say which figures you'd cite and in what order—and whether you'd include all three.

One good approach

Opening (the finding, BLUF): "Our nightly performance degradation is caused by the 8 p.m. reporting batch job, and rescheduling it to off-peak hours should eliminate the spike." Evidence order: cite A first (here's the nightly spike—Figure A), then B (here's that it lines up with the batch job—Figure B). That's the argument: symptom, then cause. C (the full schedule table) probably stays out of the body—it's reference detail leadership doesn't need to see the point; put it in an appendix if anyone wants to verify. You led with the finding, ordered the figures to build the case (effect → cause), and cut the exhibit that didn't serve the argument. That's interpretation, not presentation.

9.8 The Challenger Charts: When a Visual Fails to Make Its Point

We return, as promised, to the cold January night—this time as a problem of data display. We keep strictly to the verifiable historical record; the facts need no embellishment.

The engineers at Morton Thiokol, the contractor for the Space Shuttle's solid rocket boosters, were concerned the night before the Challenger launch that the cold forecast would compromise the O-rings—the rubber seals in the booster joints. They had data from prior flights linking cold temperature to O-ring damage. They prepared charts and faxed them to NASA, and argued their case on a late-night teleconference. They did not persuade. The launch proceeded the next morning in record cold; an O-ring failed; the shuttle was lost and seven crew members died.

Edward Tufte later analyzed the actual charts the engineers used, in his work on information display (Tier 1: Tufte, Visual Explanations, 1997). His critique is a data-visualization lesson written in the most serious possible ink. The engineers had the data, but the way it was displayed did not make the one fatal relationship—colder temperature, more O-ring damage—visible and unmissable. According to Tufte's analysis, the temperature-and-damage information was spread across multiple charts and tables, the data was ordered in ways that didn't foreground the trend (for instance, not simply sorted by temperature so the pattern would leap out), and the single most important comparison—damage as a function of temperature, with the forecast cold far outside all prior experience—was never isolated into one clean graphic that showed it at a glance. The decision-makers, reading under time pressure, did not assemble the buried relationship from the scattered evidence.

What would a Tufte-clean display have looked like? A single scatterplot: temperature on the x-axis, O-ring damage on the y-axis, every prior flight as one point, sorted naturally by the axis so the eye sees damage climbing as temperature falls—and then the forecast launch temperature marked far to the cold end, outside the entire range of prior experience. One figure. One relationship. The point unmissable: we have never flown this cold, and cold is exactly when the seals fail. The data for that figure existed. The figure did not.

⚠️ Warning "The data was in there" is not the same as "the reader saw the point." The most dangerous data-display failure is the one where every necessary number is technically present—across enough charts and tables—but no single figure isolates the one relationship that matters, so the reader, scanning under pressure, never sees it. Accuracy is not the bar. Making the critical pattern unmissable is the bar. When the stakes are high, ask not "did I include the data?" but "does one clean figure make the decisive relationship impossible to miss?"

This is the same lesson as Chapter 4's structural reading of Challenger, now at the level of the chart: there, the buried conclusion; here, the buried relationship in the data. Both are failures of making the point unmissable. And it connects forward to Chapter 38 (ethics), where we treat the full weight of the case—because when your figures inform a high-stakes decision, designing them to make the truth visible is not just craft. It is responsibility. A chart that hides a fatal trend in decoration or scatter has not merely committed a style error; it has, in the situations that matter most, failed the people who depended on it.

We rebuild a clean version of exactly this kind of temperature-versus-damage display, step by step, in Case Study 2.

🔄 Check Your Understanding In one sentence, what was the data-display failure in the Challenger charts (as distinct from the structural failure of Chapter 4)?

Answer

The decisive relationship—colder temperature produces more O-ring damage, and the forecast was colder than any prior flight—was spread across multiple charts and tables and never isolated into a single clean figure (e.g., a temperature-vs-damage scatterplot sorted by temperature with the forecast marked outside all prior experience), so the pattern that should have leapt out at a glance stayed invisible to time-pressured readers. The data was present; the visualization failed to make the pattern unmissable.

📐 Project Checkpoint

By now your portfolio's centerpiece—the technical report (piece 1 of 7)—is clear at the sentence level (Chapter 3) and organized top-down with informative headers and a stand-alone summary (Chapter 4). This chapter adds the report's visual evidence and, more importantly, the writing around it. You'll do three things.

1. Place at least one figure or table—and justify the choice. Find the place in your report where you have quantitative or comparative information (results, a benchmark, a trend, a breakdown). Decide deliberately: figure, table, or prose? Apply §9.1—is the point a shape (figure), an exact lookup (table), or one or two numbers (prose)? Write one sentence in your notes justifying the choice by the reader's task. If you choose a figure, sketch it clean: no chartjunk, direct labels, the right chart type for the relationship (§9.3). Because this book is text, you may describe the figure in words with full alt-text (as every figure in this chapter is described)—that itself is a skill the accessibility standard requires.

2. Write an interpretive caption. This is the core deliverable. For your figure or table, write a caption that reaches Level 3 (§9.5): it states what the data means, names the key number in words, and points to the implication—not "Figure 1: Results," but a one-sentence claim a caption-only reader could act on. Use the template if you're stuck: [finding], shown by [comparison]: [values], which means [implication].

3. Fix how you write about the data in the body. Find every sentence in your report of the form "Figure/Table N shows…" and rewrite it to lead with the finding and demote the figure to a citation (§9.6): claim + number + (Figure N). Run a quick reverse outline (Chapter 4) on your results section afterward—do the first sentences now state findings rather than announce exhibits?

Deliverable: your report now contains at least one well-designed figure or table, an interpretive (Level-3) caption, and results-section prose that leads with findings and cites figures rather than pointing at them. Keep your "before" captions and sentences—comparing them to the revised versions is part of the learning, and you'll want them for the Chapter 40 reflection.

Next chapter (Ch 10, design and layout) takes the visual thinking up a level—from the individual figure to the whole page: typography, white space, hierarchy, and color, so the document that holds your figures is itself easy to move through.

9.9 Common Mistakes and Practical Considerations

Mistake 1: The label caption. Captions that name the figure ("Figure 3: Churn by cohort") instead of interpreting it. The single most common data-display error, and the most consequential because caption-only readers get nothing. Fix: every caption reaches Level 3—say what it means, with the number and the implication.

Mistake 2: "Figure N shows…" prose. Sentences about the figure instead of about the finding. Fix: lead with the claim and the number; demote the figure to a parenthetical citation (Figure N).

Mistake 3: Chartjunk. 3-D bars, gradients, heavy gridlines, rainbow colors, redundant legends—decoration that obscures the data and sometimes distorts it. Fix: maximize data-ink, delete everything that isn't carrying information, label directly.

Mistake 4: The redundant figure. A figure that duplicates what the prose or the table already says, earning no place of its own. A two-bar "before/after" chart next to a sentence stating both numbers; a pie chart of percentages already listed in the text. Fix: each figure carries information the words don't; if the prose already says it clearly, cut the figure (theme 6—every element earns its place). Conversely, never make the reader read the same numbers in a table and a figure and the prose.

Mistake 5: The under-captioned table. Tables get even less captioning love than figures—often just "Table 2: Results." A table needs a caption that says what to notice in it, plus clear column headers with units, and ideally the key cell highlighted or called out in the prose. A table the reader has to mine is a table the reader skips.

Mistake 6: Misleading axes. The most ethically fraught error. A truncated bar-chart axis (starting at 90 instead of 0) makes a tiny difference look huge, because a bar encodes value by length. A dual y-axis can manufacture a correlation by scaling two unrelated series to overlap. Inconsistent or reversed scales mislead. Fix: start bars at zero; be cautious with dual axes; label scales honestly; if you truncate a line-chart axis to show a trend, say so. (This is where data display becomes an ethics issue—Chapter 38.)

Mistake 7: Describing the chart's geometry instead of its meaning. "The line rises, then dips, then rises again." The reader can see the line; tell them what it means ("growth stalled, then resumed after the redesign").

It depends — a few honest nuances. Field conventions vary: some journals require a "descriptive" caption style (stating what the figure is without interpretation), with the interpretation reserved for the Results/Discussion text; know your venue (Chapter 35). Even then, make the prose interpretive and the figure clean. Exploratory contexts differ: in a notebook, a working doc, or a peer technical review, a dense exploratory chart with a terse caption is fine—the audience has context (Dana's version for Priya, the peer, vs. Renée, the VP). The rules here are for explanatory graphics aimed at a reader who wasn't in your head. Accessibility is not optional: every figure needs alt-text describing what it shows and what it means (this chapter models that—each figure is described in words), color must not be the only channel carrying meaning (use labels, patterns, or position too, for color-blind readers and grayscale printing), and a table is often more accessible than a figure to screen-reader users, so prefer it when exact values are the point. We treat accessible document design fully in Chapter 10.

A note for multilingual writers. Captions and data sentences are forgiving territory for non-native writers, because the formula is explicit and short. An interpretive caption follows a fixed pattern (finding → number → implication) and is one sentence; a data sentence follows claim + number + (Figure N). You don't need idiom or flourish—you need the finding stated plainly with the right number. A clean figure with a plain, correct, interpretive caption reads as completely professional regardless of the writer's first language. As with structure (Chapter 4), the rules here reward learnable discipline over native fluency.

Frequently Asked Questions

How do I write a figure caption that interprets the data instead of just labeling it?

State what the data means, not just what it is. A label says "Figure 3: Churn by cohort"; an interpretive caption says "Figure 3. Onboarding speed, not price, drives churn: customers who reach first value within a week cancel at 3% versus 22% for those taking over three weeks—a 7× gap pointing to the first-week experience as the fix." Use the template [finding], shown by [comparison]: [the values], which means [the implication]. The test: a reader who sees only your caption and the chart should leave with your actual conclusion and know what to do about it.

When should I use a figure versus a table versus prose?

Match the tool to the reader's task. Use a figure when the point is a shape—a trend, correlation, or comparison the eye should grasp at a glance. Use a table when the reader needs exact values to look up or compare precisely. Use prose when you have only one or two numbers, or when the relationship between numbers is an argument that sentences carry better than any chart. And never deliver the same fact in two of them—that redundancy wastes the reader's time. Ask: will the reader see a shape, look up a value, or follow an argument?

What are data-ink and chartjunk?

They're Edward Tufte's two core ideas. Data-ink is the ink that represents actual data (the line, the bars, the points); the data-ink ratio principle says maximize that share and erase everything that isn't carrying information. Chartjunk is the visual noise that adds no information and often distorts—3-D effects, heavy gridlines, gradients, gratuitous color, decorative fills. The practical upshot: delete or fade gridlines, never use 3-D for 2-D data, label directly instead of using legends, and pick the chart type that matches the relationship. A clean chart carries more meaning with less ink.

How do I write about a chart in a report without saying "Figure 3 shows…"?

Lead with the finding and turn the figure into a citation. Instead of "Figure 3 shows the relationship between onboarding and churn," write "Customers who take over three weeks to reach first value churn 7× as often as those who get there within a week (Figure 3)." The formula is claim + key number + (Figure N): state the conclusion as the main clause, include the load-bearing number in words, and cite the figure parenthetically as evidence—just as you'd cite a source. The figure supports the sentence; it isn't the subject of it.

How do I make a chart accessible?

Three things. First, write alt-text that describes both what the figure shows and what it means (this chapter models that—every figure is described in words). Second, don't let color be the only channel carrying meaning; add direct labels, patterns, or position so color-blind readers and grayscale printouts still work. Third, when exact values are the point, consider a table instead of (or alongside) a figure—screen-reader users navigate tables well, and a table preserves the precise numbers a chart only approximates. Accessible data display is just good data display with the edge cases handled; we cover document-level accessibility in Chapter 10.

Chapter Summary

Key Takeaways

• A figure does not speak for itself. (The threshold concept.) Data is not a conclusion; the caption and the prose make a figure mean something. Caption-only readers get only what the caption says—so make it say the finding.
• Choose the tool by the task: figure for a shape (trend, correlation, comparison), table for exact lookup, prose for one or two numbers or an argument. Never deliver the same fact two ways.
• Design clean (Tufte): maximize data-ink, delete chartjunk (3-D, gridlines, gradients, redundant legends), label directly, start bars at zero, pick the chart type that fits the relationship.
• Captions interpret, they don't label. Climb from what it is → what it shows → what it means. Reach Level 3: claim + number + implication, self-contained.
• Writing about data leads with the finding, not the figure: claim + number + (Figure N), not "Figure N shows…". Interpret the number; don't narrate the chart's wiggles.
• Interpret, don't present. State the finding, then marshal the figures as evidence in an order that builds the case to an action. Cut exhibits that don't serve the argument.

Action Items

1. For every figure, write a Level-3 caption (what it means, with the number and the implication).
2. Find every "Figure N shows…" sentence and rewrite it to lead with the finding and cite the figure parenthetically.
3. Strip chartjunk from your charts: kill 3-D, fade gridlines, label directly, start bars at zero.
4. Decide figure/table/prose by the reader's task—and delete any redundant duplicate.
5. Add alt-text and ensure color isn't the only thing carrying meaning.

Common Mistakes

Label captions ("Figure 3: Churn by cohort") instead of interpretive ones · "Figure N shows…" prose · chartjunk · redundant figures · under-captioned tables · misleading (truncated/dual) axes · narrating geometry instead of meaning.

Decision Framework: how should I present this data?

Spaced Review

A few questions reaching back, to strengthen retention.

1. (From Chapter 8 — paragraphs) The given-new contract says a sentence should open with information the reader already has and end with what's new. How does that principle apply to writing a data sentence like "Customers who take over three weeks to reach first value churn 7× as often (Figure 3)"—what's the "given" and what's the "new," and why does that ordering help?
2. (From Chapter 4 — structure) You learned the inverted pyramid and BLUF for whole documents. Restate the §9.6 rule for writing about data ("lead with the finding, demote the figure to a citation") as an application of BLUF at the sentence scale. What's the "bottom line" of a data sentence?
3. (Bridging — Ch 2 audience + this chapter) Dana presents the same churn finding to Renée (VP, non-technical) and to Priya (peer data scientist). Using audience analysis (Ch 2) plus this chapter, how should the figure and its caption differ between the two—the chart Dana shows Renée versus the one she shows Priya?

What's Next

You can now make a figure carry its point and write the words that make it land. But a figure lives on a page, surrounded by text, headers, and white space—and a brilliant chart on a cluttered, airless page still repels the reader. Chapter 10: Design and Layout zooms out from the single visual to the whole document as a visual object: typography (serif vs. sans-serif, size hierarchy, line length), white space as an active design element, header hierarchy you can see, and the responsible use of color and emphasis—including the accessibility requirements (contrast, color-blind-safe palettes, screen-reader structure) that this chapter started touching with alt-text. Where Chapter 9 made your data visible, Chapter 10 makes your whole document inviting.

Practice: Exercises · Quiz Go deeper: Case Study · Case Study 2 Review: Key Takeaways · Further Reading