E-commerce AI is needed — but a chatbot alone won’t cut it

Picture this — you’re a customer at an H&M looking for pants. You enter the store — do you make a beeline for a sales assistant? Or do you make a beeline to “Pants,” sift through, find a pair you like, and only go to the assistant when you need a different color or size?

Online behavior often mimics offline behavior. Product discovery and e-shopping will not take place through a chatbot. Customers want to be helped while they’re discovering what's available; they rarely want to reach out for help while browsing. Chatbots will work for queries like “How do I cancel my account?” or “I need to return my purchase,” but not for customers who are just scrolling (or window shopping).

However, this doesn't mean AI can’t help with product discovery. A couple of months ago, I went through a deep rabbit hole on why search needs to be better for e-commerce. Product Discovery as it exists today is broken and doesn’t enable customers to ideate and discover what they’re looking for. It requires either exact-match search queries (“Skirt” or “Trousers”) or that the customers work with the categories on the site and then filter down through arbitrary filters to end up with results that are still never-ending and require hours to scroll through.

This customer experience forces customers to ideate outside the e-commerce website. And if you’re an e-commerce website, you don’t want to force your customers out. The minute they reach Google or Pinterest, they’re more likely to be led to any other e-commerce website instead of coming back to you.

My favorite online shopping experience by far is that of Google Lens. Google Lens Shopping provides multimodal, semantic search AND the ability to continuously add to your search query.

Google Lens has nailed the “filter down” of shopping. Customers often don’t know what they want until they see something similar. I might be looking for “Midi Skirts”, and come across a “White Pleated Skirt” and think “Hey! I do want pleated skirts, but in blue instead”. Only after seeing something close to what I could want do i realise that i actually do want it. Google Lens lets you do that with the simple ability of being able to “add to your search”.

The only other way I get to experience this is through “You May Also Like” recommendations. I realized I'd developed a sub-conscious tendency to click on products I didn’t actually want, but that were in the direction of what I could want. I’d click on them just to be able to get to the product page, scroll down to “You May Also Like,” and hope that the recommendation algorithm has captured what I want. Comparing this approach with the Google Lens screenshots (above), with Google Lens I get to express what it is I want to keep and what I want changed. With recommendation algorithms, I can’t tell the machine what it is that I want changed about the original product. Here, if I was looking for a blue flowerpot, there’s only one product in the recommendation carousel that got it right.

If Product discovery layouts were on a spectrum, exact-match search and arbitrary filters would be on one extreme and Google Lens Shopping on the other.

While on my Product Discovery Deep Dive, I re-imagined what a popular E-commerce site’s (Zara) discovery flow could look like if they incorporated an AI-first UX into their customer’s user flow in a way that seamlessly brings chatbot-like capabilities out of the chatbot window to meet the customers where they are when they’re just browsing.

If I were a customer looking through Zara for “Clothes for Work,” the search would not only understand what I was looking for but also help me ideate through different options

Scrolling through, I may not be interested in “Midi Dresses with Blazers,” but when I scroll upon “Button-down Shirts with Trousers,” I realize that this is something I could work with.

But I may live in a perpetually hot geographical location. So I “refine” my query and say “sleeveless shirts.” And then, liking where the results are going, I refine it down some more with “midi skirts” instead of trousers.

Resulting in a finite range of products that I want. By scrolling through what matches my aesthetic, I’m more likely to find and purchase what I actually want. But that isn’t even the best part — now, Zara knows what I’m looking for. Zara knows that I like the “sleeveless shirt + midi skirt” combination AND that I’m on the hunt for Office outfits. There’s potential to upsell more than just Office clothes. Accessories like briefcases and Formal footwear can be recommended, too.

Through this deep dive, I spoke with companies like Etsy, Fabletics, and Good American and came to 2 big learnings.

1. The potential of better search for e-commerce could impact more than just the product discovery of material goods. It can benefit any type of marketplace too (like Fiver).
2. The only e-commerce companies that could benefit from this are those with > 50,000 SKUs (unique products). Anything less, and the customer is more likely to scroll through the entire catalog with basic filters doing the trick. For example, Good American is a company with around 10,000 SKUs. However, a lot of these are made up of similar variations (like color) of the same product. Using basic filters, the relevant set of products gets narrowed down to the few that the customer can scroll through.

Companies qualifying for the “50,000 SKU” filter have started building in-house ML teams. Urban Outfitters has one, and Gap has just begun building theirs out too. So there’s hope yet — as long as these companies don’t jump onto the “Chatbot Hype Train” and invest in building the AI into the ideal customer UX.