Next week, Adobe is rolling out ‘visual similarity recommendations’ which offer AI-powered product suggestions based on what consumers are considering purchasing. And this on-the-fly use of visual interpretation and recommendation is just the start.
Now that more people are shopping online during the pandemic, brands need to facilitate the myriad ways people hunt, browse and discover products. But it’s not so easy to do that if a shopper doesn’t quite know what she wants until she sees it. Enter AI and visual similarity.
Next week, Adobe is launching a product recommendation tool that uses machine learning and artificial intelligence to surface items that are visually similar to others. The idea is to give people an easy way to find things that have visual attributes that resemble products they already like, based on similar hues, patterns and even textures.
While Amazon, Target, eBay and others have rolled out such features, Adobe is now making this option available to smaller and mid-sized merchants who use its Magento Commerce platform.
Historically, product recommendation required consumer behavioral analysis, but now these AI-powered leaps offer the potential for a better consumer experience and a flatter playing field for smaller vendors looking to compete against better-funded competitors.
“It’s interesting because you’re matching things you can’t easily describe,” says Travis Johnson, global chief executive office at the e-commerce consultancy Podean. “People won’t type in ‘rounded’ or ‘multiple colors’ and other types of terms. And you wouldn’t have a human on the other end putting in 400 descriptors for the item. What the AI can do is that and find similar characteristics which improves the search and lessens the frustration. There is an expectation now that search will take less effort.”
Other firms that have developed visual AI technology for ecommerce include Syte and Vue.ai. Advancements in the space include visual search tools that allow people to find products based on imagery from photos they upload. For example, Amazon’s StyleSnap finds items that are similar to clothing someone in a photo is wearing, such as a pair of frayed jeans or chunky sweater.
Large inventories and fast-fashion
Think of the visual similarity recommendation tool as a digitized sales associate who knows a shop’s inventory well and can help a shopper find something with similar color, shape, size, material or style, or when “You don’t really know what you want till you see it,” says Ryan Green, senior product marketing manager at Adobe.
The visual similarity tech is well-suited for e-commerce brands with a large catalog of products – for instance, a seller that offers furniture from hundreds of chair and sofa makers or a B2B firm selling thousands of electronic parts. Fast-fashion brands that churn out new items rapidly but might not have time to set up complex merchandising rules for that eighteenth seasonal winter hoodie might also benefit, says Green. “There could be something buried, and if you didn’t write a merchandising rule, I will never see that product,” he says.
Wayfair, Forever21, Neiman Marcus and others have been leaning into AI and visual search, says Kami Kris, chief operating officer of Loop Integration, the ecommerce arm of digital consultancy Kin + Carta. Kris says the improved experience has helped improve conversion rates.
But does it work?
Visual AI or “computer vision” tends to work well to surface similar patterns and colors, such as among a large collection of floor rugs, says Kevin Gumz, vice-president of applied AI and data platforms at Kin and Carta. It might have more problems detecting similarity based on texture, he says.
It’s not clear whether product discovery will always turn up what consumers might expect. Wayfair began rolling out its visual search capabilities in 2017. The online retailer serves up ’Visually Similar Items’ on product pages. However, a recent view of a Wayfair product page featuring a teal pillow made with pleated, synthetic silk-like material turned up ’visually similar’ products including a throw pillow with musical notes printed in black and white and another with a multicolored toucan embroidered on a white background.
Not only can some AI technologies produce iffy results, brands considering adding AI-based bells and whistles should be clear on goals for the features first, says Gumz. “It really comes back to the application setting,” he says. “Will this tool actually solve for the goal that we’re trying to achieve?”
At Adobe, Green says some challenges remain before computer vision algorithms do everything eCommerce brands might hope for. For example, the Magento visual recommendation system does not distinguish between branded products, and it can’t always tell the difference between a V-neck and other neck cuts in a shirt, blouse or sweater. “We’re not quite there yet; we’re still working out some of those kinks,” says Green.
Another hurdle, says Kris, affects marketers and eCommerce merchandisers. For AI to work for product merchandisers, she says, these systems should be able to show why consumers make the decisions they make. “The better engines not only use AI, but also explain to the merchandiser why certain items are getting lifted in the search results, which allows for a collaboration of machine learning with the merchandiser’s wisdom of how things should be put together as a collection.”
The more complex deep learning systems get, the tougher it is for them to explain themselves. In the world of AI algorithms, this is a concept referred to as explainability. “The key to all these search tools technologies,” says Kris, “is really around combining the ‘black box’ of AI, with what is happening from a customer intent.”