Constellations, patent pending

The AI Taste Test · Proof

Five Studies, One Pattern

Audiences sort on what an image is doing. Five perception studies on unrelated subjects kept splitting the same way, and it wasn't on style, palette or execution.

Quick answer

Across five separate perception studies on unrelated subjects, participants kept sorting images the same way. Not by style, palette or execution quality, but by what the image was doing. American symbols split on whether they stood for something or sold something. Coke imagery split on whether the drink was in motion. Santa split on whether he was busy. Robots split on whether you could tell what kind of thing you were looking at. Spooky and scary split on whether human warmth was present. Five subjects, one sorting principle, and it's the one design reviews almost never discuss.

What the studies were

Five studies published on this site, run at different times on unrelated subject matter, with general audiences and no shared hypothesis between them. Each one is a piece of audience perception mapping: an audience responds to images instead of describing them, and the pattern shows where the group agrees, where it splits, and why.

Two of these have small samples. Twenty-two people is a directional read, not a finding you'd build a campaign on, and 57 isn't far past it. The Coke study at 467 is the one carrying real weight. The pattern below is worth reporting because it repeats across all five and across sample sizes, not because any single one of them proves it.

The pattern

In every study, the winning images were doing something, and the losing images were depicting something.

Figure 1 · Five studies, one split

Five perception studies, each splitting on what the image was doing Five rows, one per study: American, Coca-Cola, Santa, Robots, Spooky or scary. Each row shows the side that won and the side that lost. American: stands for something against sells something. Coca-Cola: the drink in motion against the brand icon at rest. Santa: busy inside a rich setting against simplified or staged. Robots: you can tell what it is against you can't. Spooky or scary: human warmth present against absent. THE STUDY WON · THE IMAGE IS DOING SOMETHING LOST · THE IMAGE IS DEPICTING American 57 participants stands for something sells something Coca-Cola 467 participants the drink in motion the brand icon at rest Santa sorted on authenticity busy, inside a rich setting simplified or staged Robots US participants you can tell what it is you can’t tell what it is Spooky or scary 22 participants human warmth present human warmth absent A different subject on every row. The same split on every row.
None of these splits ran on palette, style or execution. Every one ran on what the image was doing, and the studies had nothing else in common.

Look at how each one split.

American sorted on purpose, not palette. The dividing factor turned out to have nothing to do with aesthetic style or color. Audiences separated, with real consistency, between symbols that stand for something and symbols that are selling something. Civic, commemorative and institutional uses clustered positive: USPS stamps, Olympic graphics, MLK and JFK portraits, "Made in USA" marks. Commercial and decorative uses of the identical visual language clustered negative. Same stars, same stripes, same type. Opposite result.

Coke sorted on the product doing something. The images that produced craving were cold, wet, in-motion liquid. Condensation on a glass. A bottle at the lips. The beloved brand icons and the expensive sentimental material didn't win. People don't crave the brand, they crave the drink, and the drink has to be visibly in the act of being drinkable.

Santa sorted on behavior. Authenticity judgments came down to what he was doing, not how he was drawn. The preferred images had him engaged in purposeful activity inside a rich setting: reading lists by candlelight, tending reindeer in falling snow, preparing to descend a chimney. Two things broke it. Simplified graphic depictions that strip out the context, and scenes that feel staged, commercial or temporally out of place.

Robots sorted on legibility of category. Industrial arms read as helpful. Humanoid robots in domestic settings produced discomfort. Then a retro-styled humanoid with deliberately non-realistic features broke the pattern and read fine, which points the finding away from humanoid form and toward ambiguity. The unsettling thing was not knowing what kind of thing you were looking at.

Spooky and scary sorted on presence. Spooky carried warm colors, handcrafted texture and anthropomorphic figures: burnt orange, cream, parchment. Scary carried cold palettes, digital texture and obscured or dehumanized forms. The line between them was whether human warmth and personality survived in the image. Presence within safety on one side, absence within power on the other.

Why this keeps happening

The consistent thread is that audiences read images functionally. They're answering a question the image poses about itself, and the question is roughly "what is this for."

Designers don't work that way, and there's no reason they should. Design work happens at the level of elements and their recombination: line, form, texture, color, type, hierarchy. That's the craft, and it's why a designer can look at two pieces with similar palettes and see them as related.

The audience isn't operating at that level. They're seeing complete pictures and reading intent off them. Which is why the American study produced its cleanest result on two images with an identical visual vocabulary landing on opposite sides. The vocabulary was never the variable. Function was.

Figure 2 · Same vocabulary, opposite result

Two images built from identical stars, stripes and type, landing on opposite sides Four chips across the top: stars, stripes, vintage type, red white and blue. Lines branch down to two cards carrying the identical star-and-stripes motif: a commemorative stamp in civic use, and a decorative product label in commercial use. The stamp is marked stands for something, clustered positive. The label is marked sells something, clustered negative. THE SAME VISUAL VOCABULARY stars stripes vintage type red, white and blue A commemorative stamp civic, institutional use Stands for something clustered positive A decorative product label commercial use Sells something clustered negative The vocabulary was never the variable. Function was.
A designer sees two related pieces. The audience sees one that stands for something and one that is selling something, and sorts them apart.

This gap is the reason "make it feel more premium" fails as an instruction. It's a request about register, and register is a designer-level concept. What the audience will do with the finished piece is decide what it's for, and they'll decide that fast.

What it means for how you build a test

Three practical consequences, all of which cost nothing.

Vary function, not just style. A board of twelve images in twelve treatments of one idea tests execution. A board that holds treatment roughly constant and varies what the image is doing tests the thing the audience is sorting on. Most boards do the first by default because that's what a designer naturally assembles, which is part of what belongs on a concept board in the first place.

Figure 3 · Two ways to build a board

Two boards of twelve: one varying treatment, one varying what the image is doing Left board, what most boards test: twelve tiles, the same round object in twelve different colors, which tests execution. Right board, what the audience sorts on: twelve tiles in one treatment, the object in twelve situations, at rest, in motion, wet, held, inside a setting, cropped, beside a person, being poured, in a crowd, lit, mid-bounce, hollow, which tests function. WHAT MOST BOARDS TEST one idea, twelve treatments → tests execution WHAT THE AUDIENCE SORTS ON one treatment, twelve things it could be doing → tests function The left board is what a designer naturally assembles. The right board asks the question the audience is answering.
Twelve colors of one object can only tell you which color won. Twelve things the object could be doing tells you what the audience is actually reading.

Don't strip the context out. The Santa result named this directly: simplified graphic depictions that remove context lost, and they lost against images of the same subject. Cropping to the object is a design instinct that removes the exact information the audience is reading.

Keep the neutral pile. The Halloween study's neutral group was its own category, dark subject matter handled with intellectual restraint, occult diagrams and mythic studies, and it read as neither comforting nor threatening. That's a finding. It disappears if neutral gets folded into either side, the same way the negative half disappears when nobody asks what they hate.

What this doesn't establish

Worth being straight about the limits.

Five studies on consumer and cultural imagery is not a general law of visual perception. Two of the samples are small. None of them was designed to test the pattern described here, which is the sort of thing that makes a pattern easier to see than it should be.

What it does establish is that the sorting principle is worth checking for on any given project, because when it shows up it explains more of the result than palette or treatment does. The way to find out whether it holds for a specific brand is to build the test so it can surface, and then look.

Frequently asked questions

What do audiences respond to in visual branding?

Across these five studies, function rather than style. Whether an image is doing something, what it appears to be for, and whether human presence survives in it consistently separated positive from negative responses more cleanly than palette or treatment did.

Do people judge design by aesthetics?

Less than design teams assume. In the American study, images sharing an identical visual vocabulary landed on opposite sides depending on whether they stood for something or sold something.

How many participants does a perception study need?

More than 22. That sample gives a directional read worth reporting as directional. The 467-participant study is the one in this set carrying real weight, and the pattern is credible because it repeats across all five rather than because any single one is conclusive.

What is the practical takeaway for a concept board?

Vary what the images are doing, keep their context intact rather than cropping to the object, and preserve the neutral responses as their own category.