The costliest dashboards mistakes, explained

These are craft mistakes: the specific choices on a page that make a technically correct dashboard mislead the person reading it. Every one of them is made by competent people, usually because the tool's default did it and nobody overruled it.

Nine of them, with what the reader actually sees and the smaller decision that prevents it.

What to take away

1. Building the artifact that was named in the request

Someone asks for a dashboard of support volume, and they get one. What they wanted was to know whether to move two people onto weekend cover.

The result is a page that is accurate and unusable, because it answers a topic rather than a choice. Ask what the reader will do differently, and build the smallest thing that supports that.

2. Showing the current period as if it were finished

The current month sits in the same bar chart as the previous twelve, at a third of their height, and everyone reads a collapse.

Either exclude the incomplete period, or mark it visibly and label it as partial. If the reader has to know a convention to avoid being misled, the design has failed.

3. A percentage with no base beside it

A tile reads eighty percent. It is eight out of ten, and it moves twenty points every week for reasons that have nothing to do with performance.

Show the numerator and denominator wherever a rate appears, and suppress the rate entirely below a stated base. Small groups will otherwise generate the most attention grabbing figure on the page in perpetuity.

4. Two series, two axes, one chart

Dual axis charts imply a relationship that the author created by choosing the scales. Slide one axis and the lines cross somewhere else entirely, which is why the shape appears in most catalogs of the misleading graph.

Use two charts stacked with a shared time axis, or index both series to a common starting point. Both are honest; the dual axis is a picture of a decision about scaling.

5. Defaults set for the builder's convenience

The date filter opens on last quarter because that is what was being tested. The region filter defaults to the one with clean data.

Readers rarely change a default and many never notice one exists.

Set each default for the most common reader on the most common day, and show active filters as text on the page. Put a short house standard for those defaults with the rest of your analytics foundations work.

6. Color that means different things on different tiles

Fix a small palette with stated meanings at the page level. Reserve one strong color for attention and never spend it on a category, and pair every color with a label or a shape, because color alone is not a reliable signal for every reader.

7. Sorting alphabetically

The tool sorts by name, so the reader scans fifteen rows looking for the biggest one. They will find the biggest, and they will miss the second and third.

Sort by the value the reader came for, largest first, and keep the sort stable enough that people can remember roughly where things sit.

8. Numbers the reader has to combine

Two tiles show this month and last month. The reader subtracts them. Another pair shows a count and a total, and the reader divides.

Every mental arithmetic step is an opportunity to be wrong and a reason to stop reading. Show the difference, the change, or the rate directly, and put the components underneath if anyone needs them.

9. A number with nothing to compare it to

A tile shows a figure. The reader has no idea whether it is good, and supplies a baseline from memory, which is usually wrong and always undocumented.

Every headline number carries a comparison: to the prior period, to plan, to the same period last year, or to an agreed expected range. Choose deliberately, because the choice of comparison is most of the message.

Where these come from

Seven of the nine are the tool's default behavior. That is worth saying plainly, because it changes how you fix them: the answer is a short house standard that says what your team overrules, not a design review on each page.

Write down the five defaults you always change. New dashboards then start correct rather than being corrected, and the review has something concrete to check against.

Related reading on this site

Product-level decisions behind a page, and when to retire one, are in dashboard design. Mistakes three and nine are measure properties, covered in metric shapes and traps.

For the reader's side of those problems, see questions to ask a chart. Fixed window artifacts mistaken for dashboards: see reporting. Why requests arrive shaped like mistake one: see business intelligence.

Common questions

Our tool cannot suppress a rate below a minimum base. What then?

Show the counts instead of the rate for those rows, or add the base as a second line on the tile. The point is that the reader sees the size of the group without asking.

Is a truncated axis always wrong?

No. For a measure that lives in a narrow band, starting at zero hides everything. Truncate when you must, label it clearly, and never truncate a bar chart, where the length of the bar is the entire encoding.

How do we stop old dashboards from carrying these?

Fix them when someone next asks a question about the page. Sweeping every existing dashboard at once is a project nobody finishes; fixing on contact takes care of everything that matters within a couple of quarters.

Who decides the house defaults?

Whoever builds the most dashboards, written down in an afternoon and revised when someone disagrees. It does not need a committee, it needs to exist.