Can synthetic respondents test product images and packaging?

Check what reaches the model

Ask the supplier to demonstrate the exact route from your uploaded artwork to the research output. Does the model receive image pixels, extracted text, a generated caption, or only the written concept? Which crops, page views, resolution and image transformations reach it? Keep the original files and the processed versions linked to each result.

Compatibility belongs to the named product and configuration. In the documentation checked on October 29, 2026, Qualtrics says its synthetic panels do not understand embedded images, even though they can complete surveys containing them. Ipsos describes its InnoTest synthetic workflow as text-stimulus based, with limited reactions to pictures. These are those providers' stated boundaries, not a verdict on every synthetic system or a claim about Lewsearch's image support.

Before commissioning a visual study, request:

• The supported file types, image count, resolution and processing steps for the actual research workflow, not just the underlying model.
• A record of any captions, extracted label text, filenames or other context supplied alongside the artwork.
• A way to identify unreadable, missing or failed images separately from substantive answers.
• A demonstration using your intended stimulus format, such as an isolated pack front, an online thumbnail or a shelf scene.

If only supplied concept copy reaches the system, describe the exercise as a response to that copy, not a test of the artwork. If images are converted into captions or features, disclose that conversion and evaluate whether it preserves the visual properties relevant to your question.

Separate visual recognition from consumer response

Use a small input check before asking for preferences. Can the workflow identify a feature you can verify, such as which of two packs carries a refill symbol? Check a few relevant features without supplying the answer in the question. Keep this diagnostic separate from the main evaluation so it does not prime the later response.

Then test the intended consumer outcome. Correctly identifying a symbol does not establish whether shoppers understand it, prefer the pack or notice it on a shelf. A generated explanation can suggest questions to investigate; it is not a record of someone's visual experience.

Our proposed diagnostic is to hold written context constant while changing the relevant visual feature. Also inspect whether filenames or captions reveal the preferred design. Document the full conditions and results. A response change can flag sensitivity, but does not by itself show the system responds as people do. An unchanged response is inconclusive: the visual change may genuinely be irrelevant, or the workflow may have missed it.

Match the evidence to the packaging decision

Choose the outcome before running the study:

• Label comprehension: evaluate what target consumers understand from the displayed pack, including incorrect interpretations and uncertainty.
• Design preference: compare human and synthetic preferences under the intended viewing conditions. Preference is not an observed purchase.
• Shelf findability or attention: distinguish predicted attention from measured human gaze or a timed finding task. Tobii describes eye tracking as measuring where people look and how their eyes move. A model-generated heat map is not a recording of those eye movements.
• Opening, handling, texture or taste: an image-based response is not a physical product-use observation. Include the relevant human product test when that experience drives the decision.

Reserve human evaluation results from tuning and run selection. Match the artwork, surrounding alternatives and question to the intended use, and document material differences between the model input and human presentation. For a shelf task, do not silently replace a cluttered shelf with a cropped hero image. For an online listing, evaluate the thumbnail and available zoom rather than assuming a full-resolution pack front answers the same question.

A practical brief for two cartons

Hypothetical example, not a client result: a team has two carton fronts with identical product claims, but different claim placement. Its question is whether shoppers can find the refill instruction in an online thumbnail.

A model that receives only identical extracted copy cannot establish the effect of placement from that copy. A workflow using the images or image-derived layout information can first be checked for the relevant region, then evaluated against a reserved human finding task using the same thumbnail conditions. Compare predictions of human success or finding time with observed human results, not with the model's response latency. Neither a fluent explanation nor a high stated purchase-intent score answers the original findability question.

Keep both cartons in contention if the visual input is unverified or the relevant human comparison is missing. The next useful step is to resolve that gap, not generate a larger table of preferences.

Put the boundary in the research brief

Request the exact stimuli and processing record, the outcome and viewing conditions, the separate human comparison, all planned results and the decision rule. State which questions the study will leave unanswered. This is a buyer checklist, not a certification or universal passing threshold.

For a Lewsearch study, book a demo to establish stimulus compatibility and scope before work begins. Do not assume image ingestion from a text-concept offering. The concept-testing overview explains the stated-preference screen; the blind-validation guide covers the wider evaluation.