Dashboard Design
Date: 2026-08-17
Designing for a decision rather than for coverage. Most dashboards fail the same way — they answer “what are all the numbers” instead of “what should I do”, which is why they get built enthusiastically, checked for a fortnight, and then never opened again.
Dashboard design is choosing which metrics a recurring view shows, in what order and at what grain, for a named reader and the decision they use it for.
Start from the decision
The question that determines everything else: who opens this, how often, and what do they do differently depending on what it says?
✗ "a dashboard for the ecommerce team"
→ 40 tiles, no hierarchy, nothing actionable, opened weekly by nobody
✓ "the merchandising manager checks this every morning to decide
which products to promote today"
→ one ranked table, three columns, sorted by the thing that
decides the action
If the answer to “what would you do differently” is nothing, the dashboard shouldn’t exist. That’s the most valuable outcome of asking, and it’s why the conversation is worth having before any building.
The three types, which need different designs
| Monitoring | Diagnostic | Exploratory | |
|---|---|---|---|
| Question | Is anything wrong? | Why did that happen? | What’s going on here? |
| Checked | Daily, glanced at | When something’s odd | Occasionally, at length |
| Density | Low — 5–7 numbers | High — breakdowns everywhere | Whatever’s needed |
| Design | Big numbers, deviation from expected, red/green | Segments, funnels, comparisons | Filters, pivots, raw access |
| Failure | Too many tiles → nothing stands out | Not enough dimensions → dead end | Confused for the first two |
Most bad dashboards are all three at once. A monitoring view with forty tiles has no signal; an exploratory tool used as a daily check wastes ten minutes a day. Build them separately and link between them — the monitoring view’s job is to hand off to the diagnostic one.
Every number needs a comparison
A number alone is not information. 4.2% conversion is meaningless; 4.2% against 5.1% last week is a decision.
✗ Conversion rate
4.2%
✓ Conversion rate
4.2% ▼ 0.9pp vs same day last week
── ── ── ▁▃▅▇▅▃▁ ── 14-day trend
expected range 4.8–5.4%
The comparison must be like for like — same weekday, same period length, seasonally comparable. Week-on-week beats day-on-day for anything with a weekly cycle, which is nearly everything in commerce — Seasonality.
Show the expected range, not just the number. It’s what converts “is this bad?” from a judgement into a reading, and it stops people reacting to ordinary variance — Anomaly Detection.
Things that quietly ruin them
- Percentages without denominators. “Conversion up 40%” on 12 sessions is noise. Show the base, always, and grey out or suppress cells below a minimum sample
- Averages hiding the distribution. Mean order value with a long tail describes nobody. Median and a percentile alongside it — Percentiles and Quantiles, Skewed and Heavy-Tailed Distributions
- No uncertainty. A rate from 200 sessions has an interval of several percentage points, and showing it to two decimal places implies precision that isn’t there — Communicating Uncertainty
- Truncated y-axes, which make a 2% move look like a collapse. Zero-based for magnitudes; if you truncate for a rate, say so on the axis
- Colour as the only encoding. Around 1 in 12 men has some form of colour vision deficiency, so red/green alone is unreadable for a chunk of any audience — pair it with an arrow, a sign or a shape — Colour Contrast, Inclusive Design
- No “data as of” timestamp, so nobody knows whether they’re looking at today or at a stalled pipeline — Data Quality Monitoring
- Undefined metrics. Two dashboards showing different “conversion rate” figures destroys trust in both, permanently — Metric Design
Layout that reflects priority
┌──────────────────────────────────────────────────────┐
│ ORDERS REVENUE CONV RATE AOV │ ← the 3–5 that
│ 1,247 ▼4% £48,203 ▼2% 4.2% ▼0.9pp £38.6 │ decide "is
│ │ anything wrong"
├──────────────────────────────────────────────────────┤
│ trend, 30 days, with the annotation layer │ ← context for
│ ▁▂▃▅▇▆▅▃▂▃▅▇█▇▅▃▂▁▂▃▅ ↑ price change 12 Aug │ the top row
├──────────────────────────────────────────────────────┤
│ by device by channel by category │ ← the first cut
│ ───────── ────────── ─────────── │ of "where"
└──────────────────────────────────────────────────────┘
↓ links out to the diagnostic dashboards
Top-left is the most valuable space and should hold the number that would make someone act. Everything below it exists to explain the row above.
The annotation layer is worth more than any chart type. Deploys, campaigns, price changes and outages marked on the timeline turn “why did it spike” from an investigation into a glance — Annotation and Change Logs.
Maintenance
- Instrument the dashboards themselves. Most tools report views; anything unopened for a quarter should be deleted rather than maintained
- Name an owner. Unowned dashboards drift, break silently, and keep being used
- Delete aggressively. Ten trusted dashboards beat sixty, and the sixty are how conflicting numbers enter circulation
- Version the metric definitions, linked from the tile, so “what does this actually count” is answerable without asking someone
Where it interacts
- Self-Serve Analytics — dashboards are the curated layer; self-serve is what people do when it doesn’t answer their question
- Vanity Metrics — they survive because there’s space on the dashboard for them
- Alerting on Metrics — anything checked daily to see if it broke should be an alert instead; a dashboard is a poor monitoring device because it needs someone to look
- Communicating Uncertainty — the presentation problem that dashboards get most consistently wrong