Journal of Marketing Research Scholarly Insights are produced in partnership with the AMA Doctoral Students SIG – a shared interest network for Marketing PhD students across the world.
A 2024 Bloomberg study analyzed the secondary market for luxury watches and found something remarkable: changing a Rolex bezel from black to blue (the so-called “Batman” vs. “Pepsi”) can double the price. Everything else is the same, except for the color.
Picture the scene at Rolex headquarters in Geneva: designers deliberating over whether a new Submariner variant should feature a sunburst blue dial or slate gray. Gold hands or rhodium? Should the crown be slightly larger to convey presence, or smaller for understated elegance? These choices, worth tens of thousands in market value, emerge from prototype after prototype, heritage archives, focus groups, and ultimately, the refined judgment of master craftsmen. Its artistry has been perfected over centuries. But it also relies on intuition.
An Algorithm for Decoding Intuition
In a recent Journal of Marketing Research article, researchers Ankit Sisodia, Alex Burnap, and Vineet Kumar present a deep-learning method that automatically discovers and quantifies visual design characteristics without requiring any human labeling in advance. The method uses a neural network that learns to compress product images into a small set of independent visual characteristics. It uses readily available product data, such as brand and price, as “supervisory signals” to guide the discovery of visual characteristics, ensuring the characteristics are both statistically independent and human-interpretable. Once trained, the model is also generative, capable of creating new product designs by adjusting these discovered characteristics.
Applying This Method
The researchers applied this to over 6,000 watch images. The algorithm automatically discovered six independent visual characteristics: dial color, strap color, dial size, bezel color, dial shape, and crown size. When surveyed, 86% of people agreed on what each characteristic represents, and 85% agreed with how the algorithm quantified their levels.
Using these characteristics as design knobs, they generated 729 watch variations and ran a visual conjoint study. The analysis revealed that affluent women preferred lighter dials and smaller cases, whereas less-affluent men preferred darker dials and larger cases. Using these preferences, the algorithm generated “ideal point” watches optimized for each segment. When tested, these AI-generated designs captured 16-18% choice share from existing products.
A Practical Approach
This approach requires only product images and basic data that retailers already track. The analyst does not predefine visual characteristics or specify how many to discover. It works with low-resolution images (128×128 pixels), finishes in hours, and scales across categories. The authors released their code as open source.
Practitioners can use this method to quantify visual brand equity, map competitive positioning, test design variations before manufacturing, generate prototypes, and create designs targeted to consumer segments. What took Rolex and other companies decades to refine through tradition could be decoded, quantified, and optimized by algorithms. The art remains; the guesswork is reduced.
To understand how to apply these insights in practice, we spoke with two of the authors, Ankit Sisodia and Vineet Kumar. They shared guidance on which product categories are ideal candidates, how to select the right supervisory signals, and what resources companies need for implementation.
Q: Visual design decisions are often subjective and experience-based. What practical business problems motivated you to study visual design using a machine learning approach?
A: Visual design is very important when consumers are making decisions about which product to choose. When we explore cars, prior market research shows that the biggest reason people avoid a particular car is not price, not functional characteristics, but how it looks. The practical evidence is overwhelming that visual design matters across categories, such as with companies like Apple, which initially differentiated based on design and aesthetics. We know this is very important, but it has not been studied much in the past because it is hard to quantify design and determine which characteristics are important; that’s why we started this project.
Why a machine learning approach? Prior efforts have one thing in common: the researcher has selected which parts of visual design to examine and then quantifies products on those dimensions. Our purpose was different. We asked, “What does the data say? Can we discover the different dimensions from the data without imposing our strong constraints?” The only way to do that is with a machine learning approach. It also provides scalability, as you cannot manually handle thousands of products, and it avoids subjective perceptions coloring the quantification of visual characteristics by humans. These are all important reasons we used a machine learning approach. We think it’s the only way to do this properly in a systematic way.
Q: Beyond watches and sneakers, what other product categories do you think are ideal candidates for this method? And are there categories where you would say “don’t bother,” where visual design is either too simple or too complex for this approach?
A: I think it can be applied to a lot of categories where there are product images. But one basic requirement is that the product be compositional, which means it’s composed of different parts. For example, in a watch, you have a dial and a strap. They’re all independently created and then put together. The watch is composed of many parts and factors of variation that generate it. Eyeglasses would work very well if you have the lens with a shape and a color, the frame with a shape and a color, and the length of the arm. A small number of factors can generate all the products that exist in the marketplace. That’s what we mean by compositional.
As a counterexample, paintings would be hard because a painter would create the painting altogether without you being able to figure out one particular factor of variation, and the dimensionality of a painting is unconstrained. Fashion would be somewhere in-between. A t-shirt obviously has some variation in sleeve length, number of sleeves, and number of buttons, but it could also have painting-like patterns on the t-shirt itself. Our method might identify some factors of variation, but not all factors are important to consumers. In summary, the design space needs to be constrained. It works well when you can collapse it down to a small number of dimensions that are generative over the entire set of products.
Q: One of the most interesting findings was that brand worked best for watches, but price worked best for sneakers. For a practitioner, is there any way to know which supervisory signal to use for a category before running the algorithm?
A: The method requires product images and a set of characteristics that firms typically have, such as brand, price, material, and so on. If the firm has good domain knowledge, it can form hypotheses about which signal might work. For example, if you think low-priced and high-priced sneakers differ significantly in their visual designs, then price might be a good signal for separating them. If low-priced sneakers are just imitations of high-priced sneakers in terms of visual design, price might not serve as a reliable signal. Another example: If different brands have very different visual designs, like how each watch brand has its own visual signature, where a Rolex watch is very different from a Cartier watch, then brand might be a good signal. These are heuristics the firm might use to develop hypotheses, but ultimately, they will need to test them, because even their hypotheses might not be good signals.
Think about this as leveraging human domain knowledge to establish priors over the space of exploration for signals. Managerial or industry knowledge helps narrow down which signals to test first. But we would not say there are guarantees—you have to be thoughtful about it and test multiple signals, and sometimes even combinations of signals, to find which one is best.
Q: If a brand like Nike or Adidas wanted to use your method, what non-technical steps, data, timeline, and milestones are involved? Could this be a practical managerial tool? How would marketing/design teams apply it daily?
A: We have released the source code so managers can use this method. What the method requires are product images and associated functional characteristics, such as price, brand, material, or any other relevant characteristics for the category. Once they have the images and characteristics, they can use our code with minor adjustments. They would require access to computation resources, specifically GPUs, for training deep learning models. The training is one-time and computationally intensive, but after you do that, the rest of it is not as computationally intensive.
You want to start with the business problem first. Is your product design team looking at creating a new visual design? Where do you want to explore? Do you want to stay in the region where your brand exists in visual space, or branch out to a different region? The nice thing about the method is that it’s generative, so you can explore different designs and put in business constraints. You can specify which dimensions you want to change and which you don’t. For example, Audi cars always have the four rings as a powerful visual signature in front, and they don’t want to change that, but they can change other aspects. At this point, we view this more as an aid to product designers to help amplify their capabilities, rather than as a completely automated approach.
Q: If a company adopted this approach, what team roles are essential, and how would the product design workflow change? Who would gain or lose influence in design and positioning decisions?
A: From a big picture, it empowers the company to explore hundreds or even thousands of designs that the designer might not have thought of, at least at the top of the funnel. But in terms of manufacturing, that’s still not directly affected. The manufacturing capabilities and constraints would be a very strong filter: you’ve explored 100 designs, but only 10 are actually feasible to produce at the cost you want or within the business constraints you have. So product designers could have more space for exploration, but there has to be discipline. The product management process becomes even more important because you cannot just say, go out and design some crazy design and produce it—it still has to satisfy all the business constraints.
AI can create millions of designs with no limits, but a very good designer will play both the role of editor and that of judgment, deciding which ones to pick based on their prior experience. It’s too early to say who gains or loses power specifically, but we think the role shifts toward curation and judgment rather than pure creation from scratch.
Q: You can now generate “ideal point” products optimized for segments. Do you worry about homogenization, that if every company uses this method, all products will converge toward the same optimal designs? How could firms use this approach to differentiate rather than converge?
A: We actually think the opposite. Right now, human designers are looking at what’s on the market and taking inspiration from it, exploring whether they can copy another brand or copy an aspect of another brand. But now the top of the funnel is expanding dramatically because AI lets you create so many more designs. It might actually create new brand identities and new avenues for establishing your brand in a design space. The more choices you have, the less likely we are to see similarity. Again, this is speculation, and this remains to be seen.
It’s important to see what exactly is happening in reality. If AI makes the final decision, maybe you’ll see more similarity, like how everybody’s writing is now more similar because of LLMs. But if humans have access to more AI-created designs, will they choose to differentiate more? That’s a fascinating question. We’re trying to figure this out—it’s a moving target.
Q: Can your method pick up culturally-specific design elements? For example, if a brand wants to expand into a new market, could they adapt your approach to identify which design elements will resonate with local consumers, rather than immediately hiring local designers?
A: If we train the method only on jackets sold in North America, it will be able to figure out all the characteristics of North American jackets—how the buttons work, zippers, Velcro, all of these things. It might pick up those things. But because it doesn’t see a different type of jacket closing, it won’t pick it up. It won’t be able to pick up the Chinese knot button from North American data alone.
But this raises an interesting question: can we add another category or dataset to get inspiration from? In the car market, there’s something similar. Sedans all look similar, while original SUVs looked very different from sedans; they had different frames and appearances. But when you add the two into the same model—cars and SUVs—then the model can come up with things like crossovers, which are essentially cars that look like SUVs. So we would have to think carefully about whether there’s another dataset that you can use to augment these unique cultural elements. This is a fascinating question if it can be done, and it’s very much worth exploring.
Read the Full Study for Complete Details
References
Hoffman, Andy (2024), “Is Your Rolex a ‘Pepsi’ or a ‘Batman’? It Makes All the Difference,” Bloomberg (June 11). https://www.bloomberg.com/news/articles/2024-06-11/rolex-pepsi-v-batman-buying-wrong-color-could-cost-thousands.
Sisodia, Ankit, Alex Burnap, and Vineet Kumar (2024), “Generative Interpretable Visual Design: Using Disentanglement for Visual Conjoint Analysis,” Journal of Marketing Research, 62 (3), 405–28. doi:10.1177/00222437241276736.
Go to the Journal of Marketing Research
Xiaoying Feng is a doctoral student in marketing, Syracuse University, USA.
Sichen Meng is a doctoral student in marketing, Queensland University of Technology, Australia.