Testing an AI Display Concept Before You Build It

A credible concept for a floor display used to cost a studio render or a favor from a manufacturer. So a project carried one into the meeting, two at most, and the room chose on taste. Now four directions come out of one brief in a single run. That changes the meeting. With one concept, the question was whether the team liked it. With four, the question is which one goes in front of shoppers, and most display projects have never had to answer it.

I've worked in point-of-purchase for close to four years and I build AI POP Displays. This post is about what a shopper test can learn from a generated image of a display, what it can't, and where synthetic respondents fit.

Search for this and you land in two separate worlds. One is virtual shelf testing, where shoppers look for a pack on a simulated shelf. The other is synthetic respondents, AI panels that answer surveys about product concepts and prices. Both are built around the pack and the product. Where a display does show up, in virtual store simulations, it goes in as a 3D model placed in a simulated store. Nothing I found starts from a generated concept, which is now the cheapest thing in the project.

Why does testing become the question once concepts are cheap?

Because the expensive part moves, from making concepts to choosing between them.

In AI POP Displays one brief produces up to four variants in a single run, and a second run with the brief adjusted produces more. That's more credible directions than a team can defend on taste.

And the room is often split. The retail director wants more facings, the brand director wants more brand space, and both are arguing for a shopper who isn't there. I wrote about bringing three directions into a pitch in AI display concepts for shopper marketing agencies. A test is the step after that, where the shopper gets a say.

Not every piece earns one. A shelf strip doesn't. A floor stand going into a few hundred stores, where two directions are a genuine split, does.

What can a shopper test learn from an image of a display?

Anything a shopper decides with their eyes in the first few seconds, which is most of what a display does before anyone touches it.

Whether it gets seen. Against the aisle around it, from where a shopper first meets it.

Whether the brand and the message read. The header, the claim and the product name, at the distance they'll be read from.

Whether people like it, and whether it looks like the brand. Asked of people who buy the category, not of the team that made it.

Which direction they'd stop at. The comparison the room couldn't settle.

Where the eye goes. Eye tracking works on images on a screen as well as in a store, and reports time to first fixation, dwell time and heatmaps. It shows where people looked, not why. A long look at a header can mean it's appealing or that it's confusing.

All of these are visual questions, and an image is a fair stimulus for a visual question. Pack research already runs on shelves rendered from pack images.

What has to be true of the images for the test to mean anything?

That the concepts differ only in the display.

A test compares. If concept A stands in a bright supermarket at eye level and concept B in a dim aisle seen from three quarters, shoppers pick the light, and the result reads as a preference for the design. If one concept carries the client's real pack and the other a lookalike, the test is about the pack. Same store, same light, same camera, same frame, same pack, and the display is the only thing that moves.

A general image model makes that hard, because every image is a new picture and the store, the light and the pack get decided again. Color is usually the first thing to move, as I went through in brand color accuracy in AI display renders, and in a test of whether a display gets seen, color is a large part of what's being measured.

In AI POP Displays, a comparable set is how the tool already works.

• Several directions from one brief. Up to four variants in a run.
• The real pack in every one. Product photos go in as the product, placed at scale and facing forward, so every concept carries the pack shoppers know. Which photos to send is in getting your actual product into an AI display render.
• The same store behind every concept. Change background, on the free plan, puts each one in the same setting, a supermarket, a pharmacy, a retail aisle, a boutique and others, and keeps the display, the products and the light on them.
• At the shopper's eye height. Convert to real life, on Pro, places the display in a real retail environment that matches its sector and format, shot as an in-store photograph at shopper viewpoint.
• The same camera. Generate angles, on Pro, gives a front, three-quarter, side or top view of each concept, so all of them can be shown from one angle.
• The same frame. Change ratio, on the free plan, sets every image to the same aspect, from square to 16:9 or 9:16.
• A check before it goes out. Compare puts any two versions side by side at no cost in credits, so you can confirm the set differs where it should and nowhere else.

AI POP Displays runs on Google's Nano Banana Pro. What makes four concepts comparable is the layer on top: a display taxonomy with proportions, materials that behave, retail environments as options, product load from the real packs and an edit console that holds the display while one thing changes. That's close to four years in point-of-purchase, built into the tool once.

If two directions for the same unit are splitting your team, run the brief through the free plan with the real packs and set both in the same store before anyone votes.

What can't an image test tell you about a display?

Anything shoppers do with their hands, and anything the store does with the piece.

Whether product comes off it. How far a shopper reaches, whether a pack slides out of the tray or catches on the lip, whether the top tier gets shopped or ignored.

Whether it survives the store. Whether it stands when it's half empty, whether staff can restock it, and whether it gets set up where the plan says at all.

How big it reads in a real aisle. On a screen, the size of the display is whatever the image says. A render holds proportion, not measurement, which I went through in how accurate are AI display renders.

Whether it sells. A stated preference in a survey isn't a purchase, and the gap between a picture and the real thing can be wide. In a study published in the American Economic Review in 2010, people bid 40 to 61 percent more for items put physically in front of them than for the same items shown as an image or as text.

So an image test ranks directions. It doesn't forecast a sales number. The ranking is still worth having, because it decides which concept gets the prototype money. The sales answer comes from stores, with the display in some and not in others, and the difference measured.

Where do synthetic respondents fit?

In front of the test, as a filter. Not in place of it.

A synthetic respondent is a language model answering a survey as a described consumer. A 2025 preprint by researchers at PyMC Labs and Colgate-Palmolive compared synthetic answers with 9,300 human responses across 57 surveys of personal care product concepts, and reported 90 percent of human test-retest reliability on purchase intent. The concepts were presentation slides, a text description and often an image of the product. Its authors suggest screening concepts synthetically and keeping human panels for the most promising ones.

A display's whole case is visual, and that's where the record is thinner. A preprint from September 2026 ran six preregistered visual marketing experiments through synthetic consumers, and none of the configurations reproduced more than two of the six human effects. Their answers also carried less than half the natural spread of human responses.

Two things follow for a display. A synthetic panel can only react to what it's shown, so the images still have to be a comparable set. And its place is at the top of the funnel, thinning eight directions to three before people who shop the category see them. The choice stays with the people.

How do four concepts become one piece in the store?

In an order where each step answers only what it can.

1. Generate the directions from the brief, with the real pack and the same store, as one comparable set.
2. Cut on constraints before any shopper sees them. The retailer's spec, the budget, the brand rules. A direction that can't be built for the money doesn't need a panel to rule it out.
3. Thin a long list, with a synthetic panel or an internal round, if there are more than three.
4. Test the survivors with real shoppers. In a monadic design each person sees one concept, which keeps out order effects and needs more people. In a sequential monadic design each person sees several in rotated order, which needs fewer people and lets some order effect in.
5. Prototype the winner, and check what the image couldn't: reach, load, standing, restocking.
6. Pilot it in a handful of stores against stores without it, before the full rollout.

Step one used to be the slow one, and the reason the rest rarely happened. If there's a display on your next launch plan, put its brief through AI POP Displays and bring four directions to the meeting instead of one.

Where does the prototype come from?

Step five is the one a concept image can't feed. A workshop can't build a sample from a render, because a render has no dimensions, no material thickness per part and no joints. Somebody has to redraw it first, and that's the half we're building now.

Bellto is an AI agent for brands and agencies that takes a POP campaign from brief to shop drawings. You tell it what you're launching (brand, product, channel, stores, budget, date) and it asks for what's missing. It proposes the campaign mix and materials, floor stand, glorifier, shelf strip, stopper, with the budget split per store, and designs every piece with you at real scale in parametric 3D, with proportions derived from the product and the facings. A dimension change recomputes the whole piece and re-runs the manufacturing checks in seconds. Every approved piece comes out as a 3D model, a part-by-part cutlist, dimensioned drawings, a STEP of the assembly and a cut DXF per part, a package a workshop can quote without redrawing, and it suggests manufacturers that fit by material, format and volume. The quote comes from the manufacturer.

We're building it now. The waitlist is open to any brand, and we're contacting the first ones soon to run the first real campaigns end to end. It opens by invitation, in small groups. Pricing goes first to the people on the list, and there's no card. POP manufacturers can sign up too.

For the concepts you'd put in front of shoppers, signup is free with 15 one-time credits and no card required, enough to run a brief with variants and set them in the same store. Pro is $49 a month for 150 concept generations at the founding price, with the $69 list price stated openly. Nothing you upload or render trains AI models, on any plan. When the winning direction has to become a sample, put the campaign on the Bellto waitlist.