Blog

The opportunity of visual search in retailing

Spencer Izard, research and advisory lead at Leading Edge Forum, maps out five aspects retailers need to embrace when it comes to adopting visual search.

Photo by istock.com

September 24, 2019

By Spencer Izard, research and advisory lead, Leading Edge Forum

Retailing is a visual industry and has for decades placed significant focus on the design of branding, packaging and products to attract the attention of consumers across all channels and engagement points.

Yet from the perspective of searching for a product, the consumer typically has to laboriously explore either physical or online stores in an attempt to identify an item of interest based on past purchases, a marketing piece or a memory of what someone else has that they like. A simple search either in a physical or online store can often overwhelm customers due to the amount of nearly identical products presented when searching. This issue often leads to a consumer losing interest in searching for what they want.

Reducing this issue by making it easier for consumers to find products continues to drive the demand for visual search in retailing. The increased sophistication of machine learning has seen exponential improvements in visual image processing technologies and has led to visual search maturing to a level where adoption in retailing is starting to occur.

Visual search is focused on comparing a photo taken by the consumer to the retailers' catalogue of product photos by looking for similar visual design, colors, patterns and comparative styles. From this the results derived from a “first search” can be narrowed greatly which gives consumers results more specific to their needs faster.

The potential benefits for retailers are increased customer satisfaction, reduced drop-off, increased engagement and further simplification of purchase process, all of which can lead to higher conversion rates. Visual search is currently seeing most momentum in online channels, but its true potential is within the physical store environment. Consumers often compare and contrast products across retailers so providing a mechanism for a low-friction search and comparison using visual search provides a significant improvement to the competitive dynamics of in-store retailing.

So, the question that needs to be asked is, "How should a retailer prepare for this opportunity?"

The growing importance of visual search in retailing is not an if but rather a when situation. The mainstream hype surrounding visual search is focused predominantly upon the machine learning element. However, while computer vision models are core to visual search there are other areas of complexity that must be addressed as well.

All too often retailers look for technology to reach a level of maturity and/or adoption before investing. Once a broad approach has surfaced an "arms race" often occurs between retailers to adopt a technology. This will likely occur again in regard to visual search, but to avoid being left behind retailers should consider the following now:

1.    Be clear that a computer vision model, or for that matter any machine learning algorithm, is only as capable and intelligent as the data that it is trained with.
2.    Ensure a standardized, and scalable, approach to creating and updating product images is in place to further optimize the training of computer vision models.
3.    Code product images now for use by visual search algorithms and start training models now to increase sophistication and ability to interpret.
4.    Determine the impact of visual search on the existing customer journey from search to purchase, initiated through a mobile device, wherever the customer is located.
5.    Use the assessment of visual search on the customer journey as an opportunity to further reduce points of friction that exist due to the current dynamics of text search.

Ultimately, like so many elements of the digital age, it is data that is at the heart of visual search being a success. The data a retailer creates, sources and consumes is as much a key differentiator to how they can engage with consumers to stimulate a purchase as the products themselves. Retailers, like many other industries, have struggled with data silos over the past decade limiting the true value that can be derived from data when engaging the consumer across channels. Silos of data and inconsistency in quality, and tagging, will negate any business benefit to be derived from the opportunities gained from machine learning algorithms. They require large amounts of data in a consistent form to provide the value to be gained from opportunities like visual search.

At the core of the visual search "arms race" is the differentiated value of each retailer's data sets when applied to a computer vision model. The model itself is not the unique differentiator but rather how it learns, and evolves, to understand images through the data it is provided by each retailer. What this means for retailers is that they can source computer vision models from a suitable supplier as the true value of the model is the data they own and continue to amass.

However, given the nascent nature of visual search the impact of a couple of factors have yet to be truly understood. Retailers can control the quality of what they create and feed into a computer vision model, but the consumer is a different matter. The effectiveness of visual search will be affected significantly by the quality of, or lack thereof, the images inputted by consumers into visual search. Lighting, focus and distance are a few examples of factors that will impact the quality of consumer generated images compared to the studio-quality images created by retailers.

Though it is important for computer vision models to be able to recognize products in many different forms under numerous conditions. As this technology matures retailers need to be cognizant that a minimum viable quality level of customer generated photos will be required at first as a computer vision model matures in its ability to make accurate and relevant recommendations.