Constellations, patent pending

What 500 People See · Process

Ask What They Hate

A pile of liked images is a description. The rejected pile is what turns it into a reason, and the reason is the only thing that transfers to work nobody has seen yet.

Quick answer

Most creative testing collects what an audience likes and throws away everything else. That's a mistake, because a pile of liked images tells you very little on its own. What people rejected is the foil that makes the liked pile legible. Capture both, hold them against each other, and the dividing line shows up. The dividing line is the thing you design against. This is the single cheapest change most teams can make to how they gather visual response, and it works whether or not there's a platform involved.

A liked pile is not a signal

Run the standard version of this exercise. Put a board in front of a group, ask which ones they like, write down the ones that get pointed at.

Now look at what you're holding. Twelve images somebody nodded at. They probably have things in common. They probably also have half a dozen things in common that mean nothing at all, because they were on the same board, sourced from similar places, in a similar register. You can talk yourself into any of those being the reason.

Figure 1 · One pile has no control in it

A liked pile with five shared qualities, next to liked and rejected piles with one quality left On the left, twelve liked tiles and five candidate reasons, all equally plausible. On the right, eight liked tiles above eight rejected ones with a dashed line between them; three of the candidate reasons are struck out because both piles had them, and one survives. One pile: what they liked warm palettes human faces negative space same board same source Five things in common. Which one drove it? Two piles: liked, held against rejected warm palettes human faces negative space looked like a person made it Both piles had the first three. Only one pile had the fourth.
The three struck-out reasons were true of the rejected pile too, so they can't be what the audience was sorting on. The rejected pile is what tells you that.

There's no way to tell which shared quality drove the response and which one was along for the ride. You've collected a sample with no control in it.

The rejected images are the control. They were in the same room, from the same board, seen by the same people. Whatever separates the two piles is the only variable that changed, so it's the one carrying the response.

The contrast is what produces the reason

Give someone a collection of images an audience responded well to and ask what they have in common, and you'll get a description. Warm palettes. Human faces. A lot of negative space.

Give them two collections, the responded-to and the rejected, and ask what separates them, and you get something different. You get a reason.

That shift matters more than it sounds. A description tells you what was in front of the audience. A reason tells you what the audience was sorting on, which is the only thing that transfers to work they haven't seen yet. You can't design from “they liked warm palettes.” You can design from “they sorted on whether the image looked like a person made it.”

This is also why AI is useful on the back half of a test and useless on the front half. Hand a model one collection and it will describe it fluently. Hand it two contrasting collections and ask what divides them, and the analysis gets sharp, because there's now something to divide against.

What this looks like in practice

Nothing here requires a specific tool. The mechanics are the same at any scale.

  • Capture rejection as a first-class response, not as an absence. The common failure is treating “didn't get picked” as no data. It's data. Build the exercise so a person can actively mark something as wrong for the brand, and make that as easy as marking something right.
  • Keep the neutral pile separate. Images that split the room, or that nobody felt strongly about either way, are their own category. Folding them into the rejected pile destroys the contrast you're trying to build. In the Halloween study on spooky versus scary, the neutral group turned out to be the most interesting one: dark subject matter handled with intellectual restraint, occult diagrams and mythic studies, which read as neither comforting nor threatening. That's a finding. It would have vanished if neutral had been merged into either side.
  • Test elements, not finished work. Line, form, texture, color, type. Complete logos and full layouts ask the audience to do the designer's job, and they produce a bad read because people see the whole instead of the parts and how those parts might recombine. The direction comes out of the contrast later. It doesn't go in at the front.
  • Ask the question that forces a sort. “Which of these feels most and least like us” produces two piles. “What do you think of these” produces a conversation. The first is a test. The second is a meeting.

Figure 2 · Three piles, not two

Neutral responses folded into the rejected pile, compared with liked, neutral and rejected kept as three piles On the left, a liked pile sits above a mixed pile of rejected and neutral tiles with a smeared boundary between them. On the right, three separate columns labeled Liked, Neutral and Rejected, with two sharp dividing lines and the neutral column labeled as its own finding. Two piles neutral folded into rejected the line is a smear half of “rejected” was never rejected Three piles neutral kept separate Liked what they sorted toward Neutral its own finding, not noise Rejected the control Two sharp lines instead of one blurred one. In the spooky-or-scary study the middle pile was the finding.
Merging neutral into rejected doesn't just lose the middle. It blurs the one line you ran the test to find.

Why this fixes the room, not just the data

There's a second effect, and on multi-stakeholder projects it's often the bigger one.

A room full of people looking at work has an asymmetry problem. One or two people talk through every point while the rest sit quiet. Several of them dislike what's being said and won't say so until sign-off. You lose the insight of everyone who checked out, and you never learn that the quiet majority agreed on something, because two people were working the conversation.

Capturing rejection individually and privately removes the cost of saying no. Nobody has to be the person who objects in front of the CEO. They mark the thing that's wrong, it goes into the pile, and the pile speaks for them.

That's how you end up with the outcome from the InEight rebrand: 22 stakeholders, 500 employees surveyed, one direction chosen, zero revision rounds. Not because everyone was persuaded in a meeting, but because the disagreement got surfaced before there was anything to defend.

What the rejected pile protects you from later

The practical value shows up at the end of the project rather than the start.

When someone wants to change direction late, the conversation is usually opinion against opinion, and opinion loses to seniority nine times out of ten. A documented rejection pile changes what that conversation is. The audience already told you this register was wrong for them, in numbers, before anyone spent money. Anyone who wants to move back toward it now needs evidence, not preference.

Figure 3 · The late change request, with and without the pile

Two flows for a late change request, one without a rejected pile and one with it The upper flow runs from a late change request to opinion versus opinion to seniority wins, ending in money moving back toward what the audience refused. The lower flow runs from the same request to the rejected pile answering with numbers, to a request that now needs evidence, ending in the direction holding. WITHOUT A REJECTED PILE Late change request“let’s go back to the bold one” Opinion vs. opinionnobody can prove anything Seniority winsnine times out of ten Money moves back toward what they refused WITH ONE Late change requestsame request, same meeting The pile answers500 people already said no Needs evidence nowpreference is no longer enough Direction holds and everyone knows why
Same request, same meeting. The only difference is whether the audience's no was written down when it was cheap.

That's the accountability the data buys. Everyone has agreed and they know why, because they saw why.

Frequently asked questions

Why does capturing dislikes improve creative testing?

A set of liked images shares many qualities, most of them irrelevant. The rejected set acts as a control, so whatever separates the two piles is the quality the audience was sorting on. Contrast produces a reason; a liked pile only produces a description.

Should neutral responses be grouped with negatives?

No. Neutral is its own category and often carries its own finding. Merging it into either side collapses the contrast and hides the split.

Does this work without a testing platform?

The principle does. Any exercise that lets people mark what's wrong as easily as what's right will beat one that only collects preferences. A platform matters for scale, segmentation and speed, not for the underlying logic.

What should be on the board?

Design elements rather than finished directions. Line, form, texture, color, type. Complete logos and full layouts ask the audience to do the design work and produce a weaker read.