Show Your Work

The form that informs

At Constellations we use perception testing to inform more intuitive, more effective design decisions. That work depends on a shared vocabulary, and one term we meet constantly and see defined almost never is visual data.

It is routinely confused with data visualization, and the two are not the same. Data visualization turns a dataset into a chart, a graph or an infographic to support interpretation. Visual data is information embedded directly in visual form: structured, styled and labelled so that meaning is instantly perceptible. It is not decoration. It is legibility through form.

So we put the question to a board. What does visual data actually mean to people? What associations come up, where is there alignment, and where is there ambiguity? This was the first study run on the Constellations platform, and the answer was sharper than we expected. See the full report for every image and score.

Video 1 · A deeper analysis with our founder

How the study worked

Three boards, twelve images each: dashboards, posters, packaging, renders, photography, interface screens and illustration, mixed with no grouping and no order. The prompt: drag green dots onto the images you most associate with “visual data,” red dots onto the images you least associate with it.

One hundred and eighty-one people opened the survey and 67 completed it, a 37% completion rate. At the report’s 70% threshold the 36 images sort into three groups: thirteen cleared the bar as visual data, twelve were rejected, and eleven split the room. The threshold is what turns a heatmap into a finding. Below it, an image is an opinion. Above it, the room agrees.

Meta-type is the signal

Every image that cleared the bar contains what the report names meta-type: self-referential text that documents the image’s own construction. Numerical values, parameter labels, units, dimensional annotations that describe elements inside the frame. The planet shows its surface pressure. The shadow sheet shows its blur and spread values. The lightning icon is exploded into layers, each annotated with fill, opacity and blend mode.

This is not simply text on an image. It is a feedback loop. The image shows something, the text measures it, and the measurement validates the image. The highest-scoring images do not just show data. They show their work.

Figure 2 · Every image, ranked

What won: images that show their work

Thirteen images cleared the 70% positive threshold. They range from a planet infographic to a phone dial to a printed packaging specimen to a frame of film credits, and the compositions lean modular: zoned layouts, discrete regions, clear hierarchies, dark grounds with high-contrast readouts. They look like they were made to be verified.

Medium does not matter

The array holds digital interfaces, a dashboard, a phone screen, an audio player, and it holds print. The Küche packaging grid, a specimen sheet with numbered elements, scored 76.67% alongside the dashboards. A print sheet with its parts addressed reads as data just as clearly as a screen does. What matters is the layer, not the surface it sits on.

What lost: absence, not presence

Twelve images crossed the 70% negative threshold, and they share almost nothing visually. Monochrome line art, a vibrant gradient sphere, a geometric wireframe, an illustrated wave, an atmospheric photograph, a 3D render. What unites them is not a shared aesthetic. It is the shared absence of meta-type.

None of these images contains self-referential measurement. The sunburst tree at 90.24% resistance could represent a hierarchy, but without labels it could equally be decoration. The wireframe could be a data structure, but unannotated it is a pattern. The gradient could encode something, but without parameters it is just colour.

The OKOK exception

One rejected image is covered in text. The OKOK meter reads “designed, manufactured, and produced for true emotional reading,” and it took 81.25% red. The copy makes claims and provides no values. That is the distinction in one plate: text alone is not the signal. Self-referential, quantified text is.

The twin test

The cleanest pair in the data is two round objects with structure radiating out of them. One is a planet with its atmosphere sketched around it. One is a tree inside a burst of lines. Both are centred, both are illustrated, both are quiet. One carries three measurements and the other carries none.

The difference between the top of the wall and the bottom of it is a few lines of small type. That is the whole study in two images.

The threshold zone

Eleven images landed between 31.71% and 67.65%, and the pattern is consistent: they contain some meta-type, buried under competing aesthetic elements. The data is there. It is not leading.

Near-positive

The humanoid schematic at 67.65% has technical drawing and dimension lines, partly obscured by atmospheric treatment and a dominant figure. The collage planet at 55.56% has numbered callouts and orbital notation under a vintage aesthetic that muddies the signal.

Near-negative

The fractal tree at 31.71% has the structure of a visualization and zero labelling. The Bartledan dashboard at 42.31% shows real readouts dwarfed by dramatic imagery. The weather app at 40% displays real locations and conditions, and photography owns the frame.

The contested cases suggest a ratio rule. When meta-type occupies less than roughly 40% of the visual hierarchy, opinion splits. Secondary annotation on a primarily aesthetic image does not reliably register as visual data. The layer has to be present and prominent.

The nutritional label

Consider the nutrition panel on a bottle of soda. Almost nobody reads it. If it were missing, far fewer people would drink the product. Its job is not to communicate values. Its job is to signal that the contents have been measured, regulated and deemed safe. Visual data works the same way. The meta-type layer tells the viewer: this has been subjected to scrutiny, this is accountable, this is trustworthy. The performance of rigour produces the perception of validity.

That is why presence matters more than legibility. Participants were not asking what the data says. They were asking whether anything had been measured. It is the same mechanism we found when we asked designers what “make it pop” means and what elevated design looks like: the room reads a code, and the code can be built deliberately.

The report’s own next step is a falsifiable one: take the rejected images, add a meta-type layer, and run them again. If the mechanism is right, they flip. That is what a finding should let you do.

Frequently asked questions