Data visualisation and storytelling: choosing the chart, fixing misleading ones, writing the finding

How the eye reads a chart

Some visual differences jump out before we start reading, so a chart designer decides where the reader looks first.

The session opens with a grid of digits and fifteen seconds to count the 7s. Everyone counts line by line. On the second showing the digits are identical but the 7s are coloured and everything else is grey: all eleven are found at once. That is pre-attentive processing, the handful of visual properties the brain registers before conscious reading. The deck shows five of them: position, length, colour, size and shape.

Not every encoding is read with the same precision. William Cleveland and Robert McGill ranked them in 1984 by how accurately people compare values; the deck shows a simplified version of their ranking. Positions on a shared scale come out on top, lengths close behind; angles, areas and shades of colour sit near the bottom. Florence Nightingale's 1858 rose diagram is memorable for exactly the reason it is hard to read: values carried by wedge areas.

The practical consequence is decluttering, the subject of Knaflic's chapter 3. In the deck's before-and-after, six bar colours compete, a legend sends the eye back and forth and heavy gridlines add noise. The redrawn chart sorts the bars, writes the values on them, greys out everything except the bar that matters, and gives a title saying what to notice. Every element has to earn its place: borders, heavy grids, legends that direct labels can replace, decimals nobody reads.

Back to contents.

Choosing the chart: the question decides

Name the kind of question first, comparison, trend, distribution, relationship or composition, and the chart follows from it.

Principle 2.1 of the course says it plainly: the question chooses the chart, not the software default and not habit. The deck groups business questions into five families, each with a natural chart.

A pie chart is acceptable only with two or three parts; beyond that the slices cannot be compared, for the reasons the encoding ranking gives. When the answer is a single share, a scorecard with one number is often clearer than any chart. And a question that seems to need two families is usually two questions.

Worked example: same average, different stories

Two illustrative sellers both deliver in 10 days on average. Seller X is regular; seller Y has a long tail, with 23 orders out of 100 taking 18 days or more, so nearly a quarter of its customers wait more than two and a half weeks. The mean cannot tell them apart; only the shape of the distribution can. For delivery times, review scores or basket values, the tail is frequently where the management problem lies, which is why the course insists on looking at the spread before quoting a mean.

Back to contents.

The chart clinic: misleading without lying

A chart can deceive with perfectly correct numbers, through its axis, its scale, its unit or the period it shows.

Misleading charts are rarely made by people who intend to deceive. They come from defaults nobody questioned, or from a designer who already knew what the chart was supposed to say. The clinic shows four cases with invented data, each first in its misleading form, then redrawn.

Step by step

1. Bars that do not start at zero. The length of a bar is the value, so cutting the axis falsifies what the eye reads. Line charts are another matter: with a clearly labelled axis, they can focus on the band where the data actually sit.
2. Two vertical axes. With a second axis the designer sets both scales freely, so almost any two rising series can be made to look linked. Two stacked charts sharing the time axis are honest, and two series rising together still prove no causation.
3. Counts instead of rates. A large state has more of everything, so a map of late-order counts mostly maps population. The rate shows where delivery fails, and the denominator should stay visible because small groups swing a lot.
4. A chosen time window. Three autumn months can show a surge that two years reveal as a seasonal peak returning every November. Check the longer view and the same months a year earlier, and whether a jump is a one-off event.

Worked example: a truncated bar axis

Take two average review scores, 4.1 and 4.3 (illustrative numbers). Drawn from zero, the bars differ by about 5%. With the axis cut at 4.0, the visible lengths become 0.1 and 0.3, and the second bar looks three times the first.

For a published chart, add the source and the date of the data. The class applies the checklist to two charts from the week's news, one that works and one that does not, after a five-second first look.

Back to contents.

One dashboard, one reader, one question

A dashboard that tries to show everything answers nothing; build it for one decision owner and one precise question.

A dashboard titled after a topic, such as a sales overview, has no particular point. The course asks for one reader (the decision owner of the group's brief), one question (the precise one written in Session 1) and an answer readable in a few seconds.

The reading order the deck recommends

1. Title, top left. Western readers start there, so the title states the finding.
2. Scorecards. Two or three headline numbers taken from the top of the KPI tree: the goal and its main drivers.
3. Trend. The main measure over time.
4. Breakdown. Where, or for whom, the result differs, in grey plus one highlight colour.
5. One filter. A first version needs only one; ten filters mean the dashboard has not decided what it is for.

The five-second test simulates a busy manager: someone outside the group sees the dashboard briefly, then says what it is about and what it claims. When they cannot, the repair usually lies in the title and in what occupies the upper-left corner.

Titles come in three strengths. A topic title names the subject (delivery time by category). A finding title says what to see (category D takes twice as long to deliver). An action title says what to do (review the carriers for category D first) and suits a pitch slide carrying a recommendation. One caution from the deck: a dashboard refreshed daily can contradict its own finding title the next morning, so check that the title still matches the figures.

The hands-on part uses Looker Studio: connect the cleaned Google Sheets file, add a scorecard, a time series and a drop-down control. If the time series stays empty, the date column is usually stored as text rather than as a date.

Back to contents.

The finding in three sentences

A finding becomes useful when it says what changed, why the reader should care, and which action follows, with a number in each part.

Knaflic calls the one-sentence version of a message the Big Idea. The course stretches it into three sentences, each filled from the dashboard or from an assumption stated openly. A bracket that cannot be filled from the data reveals a gap in the analysis, not a writing problem.

The deck's worked example, with illustrative figures (not an Olist finding)

1. What happened. Category D takes 21 days to reach customers, against 8 to 12 days for the five other categories. A comparison of segments works as well as a comparison of periods, provided there is a figure and a reference point.
2. Why it matters. Its customers wait about three weeks, which threatens their reviews: a risk for the Chief Operating Officer. The link between slow delivery and low reviews is written down as an assumption, to be tested in the next session.
3. What to do next. Fund a carrier review next quarter and follow category D's delivery time. It says who acts, what they do, by when, and which measure will confirm the effect.

The first sentence doubles, almost word for word, as the dashboard's finding title. The template behind it has brackets for the measure, its direction, the values, the periods and the segment, then for the cost or risk and who bears it, then for the owner, the action, the date and the indicator to track.

Back to contents.

Key takeaways

1. The question chooses the chart: say whether it is a comparison, a trend, a distribution, a relationship or a composition before drawing anything.
2. Let position and length carry the values; use colour sparingly, to point at the one thing the reader must see.
3. Look at the distribution before quoting an average; the problem often hides in the tail.
4. Most misleading charts break a simple rule: bars from zero, one axis per chart, rates rather than counts, a time window you can justify.
5. Build each dashboard for one reader and one question, with the finding in the title at the top left, and check it with the five-second test.
6. End the story with an action: who does what, by when, and which indicator will tell whether it worked.

Glossary

Pre-attentive processing

Visual properties such as colour, size or position that the eye notices before conscious reading.

Visual encoding

The way a chart turns a value into something visible: a position, a length, an angle, an area, a colour.

Decluttering

Removing every chart element that does not help the reader see the message.

Histogram

A chart of how values are spread, counting how many fall in each interval.

Scatter plot

A chart of points placed by two measures, used to see whether they move together.

Stacked bar

A bar divided into segments that show the parts of a whole.

Dual axis

A chart with two vertical scales, which lets the designer make unrelated series look linked.

Rate

A count divided by the relevant total, which makes groups of different sizes comparable.

Dashboard

A page of indicators and charts built to answer one question for one reader.

Scorecard

A dashboard element showing one headline number.

Finding title

A chart or dashboard title that states what the data show, not just the subject.

Five-second test

Showing a dashboard briefly to someone, then asking what it is about and what it says.

Big Idea

Knaflic's name for the single sentence that carries the point of a presentation.

The other notes of the course

One public summary per deck, in the order of the course.

Previous: From data to decisionsNext: SQL essentials for managers

1. Session 1From data to decisions
2. Session 2Data visualisation and storytelling
3. Session 3SQL essentials for managers
4. Session 4Predictive analytics for decision-makers
5. Session 5Generative AI, GDPR and the AI Act
6. Session 6Pitching data insights

All the notes, as a reading list

About these notes

Written by the lecturer from the slides of the deck, as a public summary: the key ideas, the formulas and one worked example. The exercises, their solutions, the tests and the grading stay in the deck, for the students of the course.

Revision: 16 October 2026.