Visualizations

An analyst spends a morning on a real piece of work. Mobile conversion has fallen and nobody knows why, so they break the data apart for an hour and eventually find it: the drop sits in one step of the checkout, on one operating system, and only since the release two weeks ago. The freeform table on the screen holds every part of that. It is correct, every number in it can be defended, and it took genuine skill to build.

Then they carry it into the eleven o'clock meeting. Somebody asks about the second row. Somebody else is reading the wrong column and arguing with a number that is not there. The product manager, who is the one person in the room who actually has to decide whether to roll the release back, looks at forty numbers in a grid and says he will take it away and read it properly afterward, which everybody present understands to mean never.

The following week the same analyst brings one line chart with two lines on it and a single sentence underneath. The decision takes about ninety seconds.

Nothing about the analysis changed in between. The table was right the first time, and the chart the second time was drawn from that same table without one number being recalculated. What changed is that the finding had been given a shape a person can absorb while half-listening, and turning an answer into that shape is a genuinely different job from finding the answer. That second job is what this part of Workspace is for.

Why the chart arrives already filled in

Drag a line chart into a panel and something slightly odd happens: it appears with your data already in it. You did not choose a dimension, you did not choose a metric, and yet there it is, plotting the right thing.

That is not cleverness on Adobe's part, and it is worth knowing what actually happened, because everything else in this section follows from it. The chart looked around, found the nearest freeform table, and started drawing that table's numbers. It has no data of its own, it never asked for any, and if you deleted the table it would have nothing to show. A visualization in Workspace is a rendering, and the table underneath is the thing being rendered.

Most of the time this arrangement is a gift, and it is worth being clear about why. It means a chart can never drift out of step with the analysis behind it, the way a number pasted into a slide drifts the moment somebody reruns the report. It means that when a picture is wrong you go and fix the table, and every chart drawing from that table corrects itself while you watch. And it means you can put a line, a bar and a headline number beside one another knowing that all three are telling the same story, because all three are reading the same rows.

It also means something less comfortable, which is that every warning in Freeform Tables now applies to charts as well, without a word of it being repeated on screen. A chart drawn from a broken-down row is showing that breakdown's numbers, scoped to whatever the parent row was. In the table there was at least an indentation hinting at it. In the chart there is nothing at all.

So what actually breaks

The rule that the chart finds the nearest table is convenient right up until the moment it is not, and then it fails in a way that produces no error message anywhere.

Think about how a project grows over a working day. You build a table to check something, then another table next to it, then a chart, then you scroll up and add a table above the whole lot because a colleague asked a different question. Later you tidy up, because half of those tables were scaffolding and nobody needs to see them. Every one of those actions is completely ordinary, and any one of them can change which table a chart is reading.

Delete the table a chart was drawing from and the chart empties. Insert a new table above a chart and the chart may quietly switch to it, keep its title, keep its colors, and start describing a different subject entirely with total confidence. Nothing on the canvas draws a line between a chart and its table, so there is no visual clue that the relationship exists, let alone that it has changed.

The fix, once you know the problem exists, is small. Every visualization has a Data Source tab in its own settings, and that tab tells you exactly which table it is reading and which rows within that table it is using. You can point it somewhere deliberately instead of leaving it to proximity. Doing that on anything you intend to keep is the difference between a project that survives being edited and one that quietly starts lying the first time somebody moves a panel.

There are two dozen shapes, and you will use about five

Open the visualizations list and it is long. Line, bar, stacked bar, horizontal bar, area, stacked area, donut, scatter, histogram, treemap, bullet, combo, Venn, map, summary number, summary change, key metric summary, and several more underneath. It looks like a set of choices to be worked through.

It is not, really. Most analysis is served by four or five of them, and the length of the list is closer to a trap than an opportunity, because browsing it encourages picking something interesting. A chart has no business being interesting. Its entire job is to make one relationship obvious faster than a sentence could, and the shapes that do that best are mostly the boring ones.

The way to choose without thinking about the catalogue at all is to say the finding out loud first, in plain words, and then take the shape that matches the verb you used.

For a first pass the working rule is short enough to remember: line for time, bar for comparison, a summary number for the headline, and stop there. That covers most of what a business report ever needs to say. It is scaffolding rather than law, and the day you can explain precisely why a scatter or a histogram answers your question better than a bar would, use it and forget the rule.

There is one shape worth being actively suspicious of, and it is the one that gets reached for most. A donut asks the reader to compare angles, and people are measurably poor at comparing angles. Try it on the picture below: the same five values are drawn twice, once as a donut and once as bars, and putting them in order is a small task on one side and no task at all on the other.

A donut also stops working entirely somewhere around four slices, and almost nobody stops at four. The horizontal bar answers the same question, holds any number of categories, keeps its labels readable, and can be understood from the back of a room. If somebody asks for a pie chart because the last deck had one, this is the substitution to make quietly.

The visualization catalogue

⤢

A long list, and most analysis is served by the four or five nearest the top. Freeform table, Cohort table, Fallout and Flow are the ones this module takes apart.

Two of them draw no data at all

Sitting in the same list, easy to scroll past, are two components that plot nothing whatsoever. Text puts a paragraph on the canvas. Section header puts a heading on it. Neither performs any analysis, neither touches a report suite, and between them they do more for a project's usefulness than any choice of chart.

The reason is worth spelling out, because the failure they prevent is so ordinary that most people have stopped noticing it. A project built over an afternoon of genuine investigation is completely legible to the person who built it. They remember which question each panel was answering, why that segment is applied, and what the odd-looking table in the middle was for. Hand the same project to a colleague a week later and none of that is present. The tables are all still correct. The charts are all still accurate. And there is nothing anywhere on the canvas that says what was being asked or what the answer turned out to be, so the colleague reads forty numbers, forms no conclusion, and closes it.

Which is how a good piece of analysis ends up being opened exactly once.

Follow along: build three pictures, then break them on purpose

The attachment between a chart and its table is easier to trust once you have watched it work and then watched it fail, and the failure is much better met in a project nobody depends on.

There is more to each of those steps, though not much more that matters. Step 7 is the one worth doing deliberately, so that the first time a chart empties on you it is a thing you recognize rather than a thing you panic about with a stakeholder waiting.

Every chart is a drawing of a table

Every visualization in Analysis Workspace is a drawing of a freeform table. It holds no data, it inherits every decision the table made including the ones you have forgotten about, it attaches itself to whichever table happens to be nearest unless you tell it otherwise, and it empties without complaint when that table goes away. The Data Source tab is the only place that relationship can be seen, and checking it before deleting anything is the difference between a project that survives being edited and one that starts describing the wrong subject.

Choosing among two dozen shapes matters far less than the catalogue suggests. Line for time, bar for comparison, a summary number for the headline, a sentence of text saying what was found. Most of what makes a project persuasive is not in the list of visualizations at all, and the components that plot nothing are doing more work than the ones that do.

Four entries in that same list are different in kind, though. They do not draw an existing table, they run an analysis of their own, and each one answers a question a freeform table cannot express however it is arranged. Cohort Analysis is the first of them, and it stops asking what happened this month in order to ask something harder: of the people who arrived in January, how many were ever seen again.