Do not pick the chart until you have the point

Friday, 16:40. The investor update goes at 17:00. Someone asks whether this month’s traction should be a pie of cities or a clustered bar of GMV, agent actives, and collections. That is the wrong question.

Write the point in one sentence first. Chart type is downstream of that sentence. If the sentence is empty, ship a table or ship nothing. You’re not missing a chart type. You’re missing a sentence.

Why these three models

The decision is which graphic ships in this week’s investor update, traction audit, or board pack. The founder decides. The clock is the send. Treat it as a design problem and you pick a template. Treat it as a data problem and you dump every series. Treat it as an argument and you name the claim, then pick the encoding that can carry it. The third one is the job. It hides Cairo’s other use of a graphic: a tool the reader can query.⁹ This week’s update is a story. The internal ops screen is a different article.

Ignored, a weak graphic keeps leaving the building. The investor prices the picture they can see, not the ledger they cannot.

Three features fire. The binding constraint is the input, not the apparatus. The reader is a detector on a noisy channel. The number only means something against a field.

That routes to data-lever-first (equilibrium), signal-threshold (random), and base-rates (cycle). Name the point before the encoding. Encode the comparison on a common scale, and refuse a chart that only restates the sentence. Put the benchmark on the chart and pair the far view with the near view. LOOP is folded: the investor relationship is repeated, but this week’s lever is a graphic they can read in ten seconds. Tal and Wansink ride inside the second card. Who draws the slide is a later article.

1. The sentence is the lever. The chart comes after

Stephanie Evergreen opens the book with the question most packs never ask. “What’s your point? Seriously, that’s the most important question to ask when creating a data visualization.”¹

The worked case in that chapter is a founder-shaped trap. A client sent a clustered bar of parents and students across several questions. The bars invited a comparison the client did not mean. The actual finding was that student expectations of college sat far below their parents’. Once they said that out loud, the headline became the point and the chart type changed: a slopegraph first, then a single large number, or a simple pie, once they decided the only number that mattered was how few students expected to go.²

The kill rule sits two pages later. Figuring out the point “reveals the best way to visualize the data.” If you do not have a point, “you probably shouldn’t bother with graphing the data.”¹

Translate that onto this week’s sheet. A last-mile founder in Kampala. Orders arrive on WhatsApp. Agents collect on mobile money. The Friday note has three series: order GMV at UGX 186 million against 179 last month, agents who completed at least one drop at 412 against 398, and the share of invoices collected inside 14 days at 71 percent against 84. Someone builds a clustered bar of all three, two months, six bars, title “Traction snapshot.” It looks finished. It hides the only finding that should have shipped. Volume was roughly flat. Collections fell 13 points. That is the sentence. The clustered bar is the parent-and-student chart again: a comparison the author did not mean, wearing the costume of a report.

Once the sentence exists, the encoding falls out. One large number for 71 percent, with last month and the 80 percent covenant on the same line. Or a slopegraph of collection rate. GMV and actives go in a two-line table. They are true. They are not the point.

The original model is older than this fold-in. Fix the input before you enlarge the apparatus. A curated subset can beat the full set. Founders invert that order every Friday. The sheet is a mess, so they pick a gallery chart. The binding constraint was the claim.

Data-lever-first, the equilibrium lens

• Assumes: one story can be named; one chart carries one message.
• Fits because: the pack is being decorated before the sentence is written.
• Breaks when: the reader is exploring, not being told, or the encoding class is genuinely the binding constraint.
• Evidence: grade B. The ordering is the model’s claim. Evergreen’s tree is a practitioner packaging of that ordering.
• Counteracts: opening the chart menu first.
• May reinforce: deleting a useful table because it does not feel like a “point.”

2. Encode the comparison, or admit you are asking them to guess

The second lens is the detector. The investor has ten seconds. They will read the picture before the paragraph. Some encodings transmit a magnitude. Some destroy it before it arrives.

Evergreen’s defence is Cleveland and McGill (1984), as she reports them. Position on a common scale sat at the top of their hierarchy: the easiest visualisation for people to interpret with accuracy.⁴ Angle produced the most errors. Pie charts produced the most errors. Area (bubbles), volume (anything 3D), and curvature (donuts and their cousins) are where people guess.⁴ She is not telling you to make every slide a bar. Story first. Then climb the hierarchy inside that story’s shortlist.⁴ For a comparison, her later chapter names the practical winner: the dot plot, dots on a line.⁵

Collections versus last month versus the covenant is a magnitude call. Put all three as positions on one shared scale. A donut of collected versus not asks the reader to judge an angle. A bubble of cities by GMV asks them to judge area. A 3D column that “makes growth pop” asks them to judge volume. Those are noisier detectors. The miss is the comparison the investor fails to read. The false alarm is the comparison they think they read.

The same detector produces a second error that looks like extra proof. Evergreen reports Tal and Wansink (2014): people believed a medication was more effective when the materials included a graph, even if the graph added no substantial information.³ Graphs, she writes, “seem to add credibility to data, even if they don’t contain any new insights beyond what already exists in the narrative.”³

That is the traction slide that restates “GMV grew 40 percent” as a giant arrow. The sentence already carried the claim. The graphic added no comparison. It raised belief anyway. A chart that restates the sentence is not a second source. Delete it, or add a comparison the sentence does not contain.

If you draw columns, start the axis at zero. Otherwise the length of the bars sends a distorted message.⁷ She allows that a line-graph axis can be truncated to support a decision. If you truncate, write why on the chart. That is a gate, not this article’s thesis.

Signal-threshold, the random lens

• Assumes: the reader judges magnitude by eye on a shared scale.
• Fits because: the update is a detector, and encoding is the channel.
• Breaks when: a one-slice pie spotlights a single number, or a line axis is truncated on purpose and labelled.
• Evidence: grade A on the detector. Encoding clause is Cleveland and McGill via Evergreen.
• Counteracts: treating a restating chart as extra evidence.
• May reinforce: making every slide a bar and calling that rigor.

3. Compared to what, and compared to whom

The third lens is the one the memo keeps stealing from the chart.

Evergreen’s opening to the benchmark chapter is the sentence most board packs fail. A standard, target, benchmark, average, national norm, or long-term goal “is too often missing from our data displays.” People know the department goal. “Sure, it’s written down in a memo somewhere.” Without it on the graph, “we limit the amount of information a reader can pull from our graphs.”⁶ Two pages later she names the first reader question: “Compared to what?”⁶ The encodings she actually names are a shared benchmark line, a combo or bullet when each category has its own target, and indicator dots that mark a problem that needs attention.⁶

Last month is a neighbour. It is not a standard. The Kampala pack that shows 71 percent against 84 percent last month still has not answered whether 71 is good, bad, or close. The 80 percent collection covenant in the facility letter is the standard. Last year’s same month is a better far view than June if GMV always dips in the rains. Put the benchmark on the chart, not in the memo.

Scott Klein’s pairing rule, quoted in Alberto Cairo’s 2017 draft dissertation, is the same demand in another dialect. Far view first: the ranked field, the average, the whole. Near view second: this town, this company. “Whenever possible, every number in your app should include a comparison to another,” neighbour, cluster, or whole.⁸

On an investor update the far view is the field the reader already carries. The near view is you. A collections number with no covenant and no peer is a near view with the field deleted. A TAM slide with no “us” is a far view with the company deleted. This piece is not the TAM article. The pairing rule still applies to the one graphic you are about to send. If the only defensible far number is internal (the covenant you signed, cash to stay open), use that and say so. A borrowed industry average from a deck you cannot rebuild is a vanity far view.

Base-rates, the cycle lens

• Assumes: an honest field exists, and a local unit the reader can stand on.
• Fits because: the pack shows this month against last month and leaves the standard in a memo.
• Breaks when: there is no honest near, or the far view is a vanity TAM.
• Evidence: grade A as a method. Far-and-near is Klein via Cairo, a heuristic.
• Counteracts: last-month-only comparison dressed as a verdict.
• May reinforce: importing a peer average from a different corridor.

GEER: the cuts that cost nothing before Friday

Cheapest and most reversible first. Ruin on this decision is a graphic that becomes the only evidence in a live raise. Cap that. Do not let a truncated column or a restating arrow be the sole exhibit for the number a term sheet will price.

1. Write the sentence. One line, before anyone opens the file. If you cannot, stop. Evergreen’s kill rule applies.
2. Delete the charts with no point. True numbers that do not change the decision stay in a table. Free.
3. Kill the restatement. If the graphic adds no comparison the sentence lacks, it is decoration wearing evidence. Delete it.
4. Redraw magnitude calls on a common scale. Dots or bars on one axis. No donut, no bubble, no 3D. An hour in the sheet you already have.
5. Put the benchmark on the same axes. Covenant, target, last year’s same month, or a sourced peer. The memo is not the chart.
6. Pair far and near. This company and the field, on one artefact. A number that cannot take a comparison does not ship.
7. One chart, one story. A second story is a second graphic or a table. Not a second series dumped onto the same axes.

RADAR: what leaves the building this week

Do now, by T+3, one sitting, effect visible on the next send. Open last week’s update. Write the point of each graphic in one sentence. Delete or redraw any graphic whose sentence is blank, or whose picture only repeats the sentence. Reversible: you can put a table back. Dominant across every scenario about whose taste the pack should please. The effect arrives when the next reader can restate the claim without you in the room.

Hedge, by T+14, premium is a template, cover live before the next board pack. Freeze four fields at the top of the file: the sentence, the story class (one number, a comparison, a benchmark, a trend), the encoding, and the comparison (neighbour, cluster, or whole). The premium is twenty minutes of friction. The cover has to be live before the next pack is assembled, or the hedge is a note to self. If a field is blank, that graphic does not ship.

Defer and trigger, size declared when the trigger is set. Hiring a visualisation contractor, or buying an investor dashboard, is spend on the view layer. Pre-commit the observable: four dated updates in a row in which a second person, given ten seconds, restates the point of each graphic without help. When that is true, you may pay for polish. Not before. A prettier channel will not fix a missing sentence.

A DO NOW whose effect arrives after the DEFER trigger fires is misordered. Do not buy the dashboard on Tuesday and write the sentences on Friday.

CHAIN: how a time-poor investor actually reads the pack

Name the shape first. Most packs overshoot on decoration and collapse the first time a reader is asked to repeat the claim. The dashboard-looking page was the overshoot. The cliff is the partner who forwards one slide and cannot say what it argued.

Run the break test before you borrow last year’s “this is how we present.” Has a rule changed, has an actor entered or left, has a measurement become a target? The last one fires often. If the chart that “pops” is the one that gets praised on Monday, you are now selecting for pop.

The class to match is one-way argument slides read by time-poor professionals, not “startups that use colour.” Diligence memos, credit notes, board packs. Readers skim. They take the visual as the argument. They ask “compared to what” whether you printed the answer or not. A sentence plus a comparison on a common scale survives that reading. A gallery of pies does not.

Subtract the counterfactual before you credit the new template. A pack that “landed better” after a redesign may have landed because the sentence was finally written. The test is whether a stranger can restate the point. Usually the colour was not the variable. Mobile-money exports and signed LPOs already timestamp the near view, which makes an honest comparison cheaper to draw.

Matrix-break flag. All three lenses assume the graphic is a story you chose. The day the pack is generated from a prompt, the sentence is the only owned claim left. Keep the sentence, the source, and the comparison in a file a model did not write.

What three lenses still cannot tell you about the slide

They cannot tell you whether the number is true. A perfect encoding of a cleaned cell is a high-fidelity lie. The journal behind the graphic is a different decision, and it has its own article.

They cannot tell you the exploratory case. Cairo’s tool, the graphic that lets a reader find their own town, is outside this boundary on purpose. Shipping a story at a reader who needed a database is a FRAME error, not an encoding error. He says any rigid taxonomy is doomed.⁹ Use the split as a purpose test.

They cannot tell you which comparison the other side already believes. An investor who thinks collections in your corridor run at 90 percent will read your 71 against that private far view, whether you printed one or not. They did one step. They will not reconstruct your covenant from a memo on page two.

One property no member models: once you require a point, people stop putting the awkward series on the page. The quiet collection-rate drop was absorbable in a six-bar forest. Named, it becomes a conversation, and some founders then omit it. That omission does not appear in any encoding hierarchy.

The one action that survives the ignorance: before this week’s send, write the point of the graphic in one sentence. If you cannot, do not graph. If you can, put the benchmark on the same axes and encode the comparison as position on a common scale. Then send.

Who signs the sentence

The person who needs this is the one who is about to ask the designer to “make the traction pop,” or the one who is about to paste last month’s clustered bar into this month’s note. The cheapest first test costs twelve minutes and no money. Write the sentence. Hand the draft to someone who has not seen the sheet. Ask them to say the point back. If they cannot, the graphic was never the argument. The sentence was missing, and the send is not ready.

Sources and notes

1. Stephanie D. H. Evergreen, Effective Data Visualization: The Right Chart for the Right Data, SAGE, 2017, Chapter 1, “Why We Visualize”, PDF pp. 20 and 23. “What’s your point? Seriously, that’s the most important question to ask when creating a data visualization.” And: “If you find you don’t have a point, you probably shouldn’t bother with graphing the data. We visualize to communicate a point.” Locators are PDF pages in the 2017 first edition (ISBN 9781506303055). The print-index offset is not a single integer, so PDF page is the citable locator. Publisher author page (identifies the book; SAGE is a known hostile host for full text): us2.sagepub.com/en-us/nam/author/stephanie-d-h-evergreen.
2. Evergreen, ibid., Chapter 1, PDF pp. 20 to 22. Worked client example: a clustered bar of parents and students hid the finding that student college expectations sat far below parents’. Naming the point changed the headline and the chart type, first to a slopegraph, then to a single large number or a simple pie once the team decided the only number that mattered was how few students expected to go.
3. Evergreen, ibid., Chapter 1, PDF p. 24, citing Tal and Wansink (2014) as she reports them. “Graphs and formulas seem to add credibility to data, even if they don’t contain any new insights beyond what already exists in the narrative.” People who read medication-efficacy materials “believed the medications were more effective when the materials included a graph, even if the graph didn’t contain substantial or additional information.” Primary-paper page numbers are not claimed here.
4. Evergreen, ibid., Chapter 1, “Which Chart Type Is Best?”, PDF pp. 27 to 29, citing Cleveland and McGill (1984) as she reports them. Position on a common scale was the easiest visualisation to interpret with accuracy. Angle produced the most errors. Pie charts produced the most errors. Area, volume, and curvature sit at the bottom of the hierarchy. “We should be striving to graph as high up in this hierarchy as possible.” She adds that story, not the hierarchy alone, chooses the type. Her source note gives the paper as Journal of the American Statistical Association 79(387). Landing for the paper she cites, title and year only: tandfonline.com/doi/abs/10.1080/01621459.1984.10478080. Primary-paper page numbers beyond her citation are not claimed.
5. Evergreen, ibid., Chapter 3, PDF p. 80. “Another new graph type, dot plots, plunks dots on a line, the easiest visualization for humans to interpret with accuracy.”
6. Evergreen, ibid., Chapter 4, PDF pp. 139 to 141. The target, benchmark, average, national norm, or long-term goal is “too often missing from our data displays” and “written down in a memo somewhere.” “Adding benchmark information increases and deepens nearly any story told in this book” and answers “Compared to what?” Encodings she names: a shared benchmark line; combo or bullet when each category has its own target; indicator dots that mark a problem that needs attention.
7. Evergreen, ibid., Chapter 1, “When Visualization is Harmful”, PDF p. 25, Figure 1.4 caption. “In a column graph, the axis should always start at 0. Otherwise the length of the bars sends a distorted message.” She notes justifiable reasons for truncating a line-graph y-axis (pointing to Chapter 9) and that “distortion is real, common, and harmful.” Pandey et al. are cited only as she cites them. Their figures are not restated here.
8. Alberto Cairo Touriño, Nerd Journalism: How Data and Digital Technology Transformed News Graphics, doctoral thesis, Universitat Oberta de Catalunya, draft dated 5 June 2017, Chapter 3, printed p. 94. Long quotation of Scott Klein (2013) on the far view and the near view, including “Whenever possible, every number in your app should include a comparison to another” (neighbour, cluster, or whole). Klein is quoted as he appears in Cairo. The 2013 original was not opened for this piece. Thesis landing: uoc.edu thesis page. The TDX handle for the same work bot-walled a fetch at writing time, so the UOC page is the reader link.
9. Cairo, ibid., Chapter 1, printed pp. 30 to 31. A data visualisation is “intended to enable exploration, rather than an unidirectional conveyance of information.” It does not necessarily tell a story or have a sequential structure. Hybrids are common, and “any attempt at devising a detailed and rigid taxonomy of news graphics is doomed.” Used only as a FRAME split: this week’s update is a story, not a tool.

A note on what is deliberately absent. This is not the operator-dashboard piece, and it is not the article on who owns the claim. Truncated axes as a standalone decision were skipped so this one could stay on the sentence, the encoding, and the missing line. Cleveland, Tal, and Pandey are cited only as Evergreen cites them.

Joshua Agonya Pi’Rwot, Founder.