Can AI synthetic shopping tell you which shelf a shopper picks?

The briefThe use caseWhat did the synthetic shopping simulation need to do?

The use case was simple:

1. Simulate POV view in a real shopper environment.
2. Swap out shelf placement, and see if consumers could or would see the items clearer, or which shelf would make it more obvious.
3. POV movements needed to simulate a real human: shifts of visual line of sight, the turning of POV while entering an aisle, etc.

WorkflowHow I did itWhich tools and inputs did the workflow use?

• 10 reference images (actual photos) of the environment
• Claude for prompting (with a Claude skill for Seedance)
• Single prompt each run
• Seedance 2.0 (for video), GPT 2.0 (for shelf configurations)
• And MULTIPLE MULTIPLE iterations of the prompt / generation

ResultsWhat worked and what didn'tHow accurate was the AI-generated shopper POV?

What worked

• Accuracy of shelves were spot on from reference (meaning you can simulate different shelf configurations).
• POV was natural and human like (visual accuracy and line of sight).
• Pace of walkthrough was realistic.

What didn't

• Environment sequence and simulation was inconsistent (even though multiple reference images were used, too many in fact).
• Directional stitching (i.e. exit left aisle, enter front aisle from side) was accurate, but reference stitching was not (i.e. turned right correctly, but wrong shelf space rendered on the right).
• Max 15s generation. Technically you could use the last frame to stitch the POV, or use the newly launched Seedance 2.5, or slow down the video slightly like I did.

011. Single shelf POV worksDoes synthetic shopper simulation actually work?

Shopper simulation works and can yield strong results, but you have to restrict it to a single shelf or aisle per run. Think of it like briefing different shoppers to each explore one shelf and observe it through their eyes.

022. On cost and efficiencyIs synthetic POV cheaper than a real shopper study?

The alternative is headcams on shoppers, live reactions, and physically re-arranging the shelf for every variant. Against that, synthetic POV still wins on cost, time, and speed.

033. Will a shopper actually reach for your product?Can Gen AI predict which product a shopper chooses?

Whether a shopper actually reaches for your product? Gen AI alone can't. A game engine with cognitive rules is required (e.g. colour contrast >23% against adjacency = predicted linger). POV shows what's visible, but not what gets chosen.

044. Multi-shelf exploration vs. resourceIs a full multi-shelf shopper journey worth the effort?

You can get the perfect end-to-end journey, but only by stitching footage, not from a single prompt. Once you factor in the iterations and credits, the effort may not be worth it. At that point, $50 a head for real shoppers and one staffer moving stock around might just be the better deal.

CodaMy POVCan synthetic shopping replace a real shopper study?

Synthetic POV restricted to single shelves is good enough to augment a real shopper study, but not yet to replace it.

Running a shelf or pack test? WhatsApp me or email me, or try our free predicted-attention heatmap.