← Field Notes
Strategy14 August 2026·11 min read·Chris Ma

BEFORE
THE
CHART.

Most bad dashboards aren't built with bad intentions — they're built with unasked questions. The perception science, honesty rules, and build sequence that change that.

Data VisualizationDashboardsChart DesignBI ToolsGA4

The failure mode in most dashboards isn’t the chart library or the color palette. It’s that the question was never written down before the building started. “Show performance” sends you in the wrong direction before you’ve opened a tool. “Did the Q3 campaign hit its CPA target” resolves most downstream design questions automatically: the chart type, what to cut, what to highlight, and what the headline should say.

Write the business question in one sentence before picking a chart type. If you can’t, the visual isn’t ready to build. This sounds obvious and is routinely skipped, which is why most bad dashboards aren’t bad because of technology or design skill. They’re bad because no one wrote down what decision the chart was supposed to inform.

Accuracy and readability are separate requirements, not substitutes for each other. A chart can be beautifully designed and still misleading. A technically correct chart can still be unreadable. Both bars have to clear.

Key Takeaways
  • Write the business question in one sentence before picking a chart type. If you cannot, the visual is not ready to build.
  • Perception accuracy follows a clear hierarchy: position on a common scale (bar chart) is the most reliable encoding. Angle (pie) is systematically misjudged. Area (bubble) is consistently underestimated by roughly 30%.
  • One high-contrast highlight colour per chart is enough. Multiple accent colours shift cognitive load from the data to decoding the colour legend.
  • AI is useful for drafting, labelling, and layout. The business judgment about what to highlight and what to cut requires knowing what decision the chart is supposed to inform.
01

WHY SOME CHART CHOICES ARE WRONG, NOT JUST UNPOPULAR

#

Chart choice isn’t a matter of style. It’s grounded in how human visual perception actually works. The relevant concept is preattentive processing — the visual judgment that happens before conscious attention kicks in. Different encodings engage this mechanism differently. Some produce accurate judgments at a glance. Others are systematically misjudged, regardless of how well-designed the chart is.

Length on a common scale — the bar chart — is the most reliable encoding available. Two bars plotted on the same axis: the difference reads immediately and accurately. This is why bar charts dominate good analytical work. Not as a default, but as a deliberate match between task and encoding.

Area is reliably underestimated. When comparing circles, humans consistently misjudge the ratio — a bubble representing twice the value of another looks like maybe 1.5×. This isn’t a fixable design problem. It’s a property of how the visual system processes area. Treemaps and bubble charts are weak tools for precise comparison, not just unfashionable ones.

Angle is the worst encoding for comparison beyond two or three categories. With more than three slices, angle differences become genuinely unreadable, and errors compound as the number of categories grows. Pie charts aren’t bad for aesthetic reasons. They’re bad because the question “how do these compare?” cannot be accurately answered by looking at angles.

VISUAL ENCODING — HOW HUMAN PERCEPTION JUDGES THESE AT A GLANCEACCURATE ✓7245ABLENGTH / POSITIONjudged correctly at a glanceMISJUDGED ✗7245AREAsize difference is reliably underestimatedUNRELIABLE ✗72%28%ANGLEworst encoding for precise comparison

The same two values — 72 and 45 — encoded three different ways. Only one reads accurately at a glance.

Practical selection guide

Comparing categoriesBar chart
Change over timeLine chart
Part-to-whole (2–3 items only)Stacked bar — not pie
DistributionHistogram or box plot
Correlation between two variablesScatter plot
Hierarchical proportionTable or sorted bar chart before reaching for a treemap
02

THE HONESTY RULES — EACH ONE DESCRIBES A REAL MISTAKE

#

These aren’t stylistic preferences. Each describes a specific, common error that changes what a chart says — silently, without requiring any design intent.

Don’t connect a line across missing data

A connected line makes an implicit claim: a value existed for every point along it. When data is missing — a tracking script went down, an integration broke, an event wasn’t captured — connecting across that gap invents a trend that was never measured. Break the line, or mark the gap explicitly. The alternative is a chart that lies by continuity.

Don’t let a partial period read as a complete one

Six days of July plotted alongside six complete prior months creates a visual collapse that isn’t happening. The bar is low because it contains six days of data, not because July underperformed. Every live dashboard that includes the current period needs one of three interventions: shade the partial bar differently, label it “in progress,” or exclude it until the period closes.

WITHOUT PARTIAL-PERIOD HANDLINGWITH PARTIAL-PERIOD HANDLINGJanFebMarAprMayJun–63%?Not a drop — 6 of 30 daysJanFebMarAprMayJun↗ In progressJune reads as a collapseJune is clearly partial data

Same data — June with 6 of 30 days collected. One version misleads; the other informs.

Don’t drop a forecast’s confidence range

Any statistical forecast carries real uncertainty the model can quantify and report. Showing only the point estimate and omitting the confidence interval removes the only part of the forecast that was honest about its own limits. A narrow band is reassuring. A wide band is information — the reader deserves to know how uncertain the projection is. Both are more useful than a single line drawn through unknowns.

FORECAST — POINT ESTIMATE ONLYFORECAST — WITH CONFIDENCE RANGEforecast →forecast →upperlowerDrops the only honest part of the forecastShows real uncertainty — narrow or wide, it is information

The confidence range isn’t decoration — it’s the honest part of the forecast.

Use color to focus attention, not to fill the legend

One accent color on a muted base does more work than five colors competing for attention. Gray context bars, one cobalt bar for what matters this period. This single change — requiring no other redesign — is the fastest upgrade available for any executive-facing dashboard. The insight becomes visible without narration.

BEFORE — SIX COLORS COMPETINGAFTER — GRAY CONTEXT, ONE ACCENTJanFebMarAprMayQ3JanFebMarAprMayQ3↑ Q3Every bar demands attentionThe eye goes straight to what matters

Same six bars. The right version guides the eye without requiring a legend.

Label directly — don’t outsource meaning to a legend

A legend forces the reader’s eye to travel between data and key — a round trip with every new data point. Labels placed on or next to the data eliminate that friction entirely. This is especially important for line charts with multiple series: a label at the end of each line costs nothing and removes a constant cognitive tax from the reader.

Remove anything that competes with the main insight

Gridlines, decorative borders, 3D effects, and unused legends are each a small tax on the reader’s attention. They accumulate. A chart that has had decorative elements stripped out isn’t “minimalist” — it’s a chart where the data has more room. Every element that doesn’t serve the one question this chart answers is in competition with it.

03

DASHBOARD-LEVEL DESIGN — BEYOND THE SINGLE CHART

#

One question per view

The most underused principle in dashboard design. Every view should answer exactly one business question, with the answer placed prominently above the fold. Multiple questions crammed into one screen shift the prioritization burden to the viewer — which slows decisions and increases the chance they leave with the wrong takeaway. The question shouldn’t be implicit in the charts. It should be stated, in plain language, before the charts begin.

Interactivity should be used, not just available

Tooltips, drill-downs, and cross-filtering are valuable when they’re actually used. Review interaction data periodically: if a filter or drill-down is never clicked, remove it. Unused interactivity adds cognitive overhead without adding analytical value — and it’s the easiest thing to accumulate without noticing, because it feels like adding capability when it’s actually adding noise.

Animation is for state change, not decoration

Motion should communicate that something changed — a filter applied, a value updated. Animation that exists for visual interest distracts from the data and creates a genuine accessibility barrier for users with motion sensitivity. If removing an animation doesn’t make the chart harder to understand, the animation shouldn’t be there.

Mobile requires rearrangement, not rescaling

Charts that simply shrink on mobile become unreadable rather than smaller. The layout needs to genuinely reflow — which means designing for it explicitly as a separate layout state, not treating it as a scaling problem to handle automatically.

04

AI IN THE VISUALIZATION WORKFLOW

#

The shift worth understanding: AI tools now handle a meaningful share of the mechanical work — cleaning data, selecting chart types, detecting anomalies, generating written summaries — from natural-language descriptions alone. Typing “show sales trends for the last six months” produces a working chart in most modern BI tools, with the AI selecting the format and often surfacing an initial insight. The genuine benefit is removing the setup cost between raw data and a first-draft visual, freeing time for the judgment calls described in sections 1 through 3.

Where AI-generated charts need real caution

SQL hallucination

An AI generating SQL from scratch can silently compute “revenue” or “conversion” using an assumption that doesn’t match your team’s agreed definition. An AI grounded in an explicit semantic layer — pre-agreed, governed metric definitions — is materially more trustworthy than one improvising from a raw prompt. A metric that seems right but isn’t damages trust in the whole dashboard faster than any single right answer can rebuild it.

Summary scope

Any AI-written chart summary should be treated as provisional until you’ve confirmed the date range and filters it used. A summary that doesn’t state its own scope can’t be verified. An unverified metric shared upward is a metric that may be wrong — and wrong numbers erode trust in the whole dashboard, not just the one number.

AI forecasts — same honesty rules apply

An AI-generated predictive overlay that drops its own confidence interval is making exactly the mistake described in Section 2 — automated. The honesty rules don’t carry an AI exemption. If the model can report uncertainty (it almost always can), the chart should show it.

05

THE BUILD SEQUENCE

#

Before opening any tool, write the business question in one sentence. Then:

01

Write the question

One sentence. If you can't write it, the visual isn't ready to build.

02

Choose chart type from the perception table

Match the encoding to the question. Not from habit, not from what looks interesting.

03

Build the honest version first

Gaps shown as gaps. Partial periods marked. Forecasts shown with their confidence range.

04

Apply the one-highlight-color rule

Gray for context. One accent for what matters this period. No other colors.

05

Cut everything that doesn't serve the question

Gridlines, decorative borders, unused legends. Each one is attention drawn away from the data.

06

Verify any AI-generated content

Metric definition. Date range and filters used. Don't trust the summary before checking its scope.

07

Show it to someone outside the room

Teams that ship readable dashboards treat a confusing chart type as a defect, not a preference. Fresh eyes catch it before deployment does.

THE HONESTY CHECKLIST

#
No line connects across a real data gap without a visible break or explicit note
No partial / in-progress period is shown as if it were complete
Any forecast shows its confidence range, not only the point estimate
Color is used to highlight the one thing that matters — not to decorate
Chart type matches how the data is actually perceived (length > area > angle, wherever precision matters)
Every view answers exactly one stated business question
Any AI-generated metric or summary has been checked against a governed definition before being shared
Recommended Reading

Edward Tufte · Graphics Press

The definitive text on data graphics. Tufte's data-ink ratio principle alone changes how you evaluate every chart you produce.

Cole Nussbaumer Knaflic · Wiley

Practical guide to turning spreadsheets into narratives that drive decisions — focused on business audiences, not academic ones.

Alberto Cairo · W. W. Norton

A journalist and visualization researcher explains the seven ways charts mislead, and how to read them critically.

Continue the conversation

If this changed how you think about it — or you think I'm wrong — I want to know.

Corrections, disagreements, and applications all welcome. Replies go directly to Chris.

Get in touch →
Field Notes · PodcastHost + Expert · Gemini TTS

BEFORE THE CHART

~6-8 min

1× · Two speakers · tap to play