Reimagining Retail: How retailers are using GenAI and when they should absolutely not use it

Audio by Sara Lebow, Blake Droesch, and Carina Perkins | UPDATED: Oct 16, 2024

Episode Transcript:

Sara Lebow (00:00):

This episode is brought to you by TikTok for Business. Are you ready to rizz up your brand? Well, you're in luck. TikTok is where discovery drives outcomes, and three in four users say they will buy from a brand they've seen on our platform. Learn more at tiktok.com/business.

(00:19):

Hello, listeners. Today is Wednesday, October 1st. Welcome to Behind the Numbers Reimagining Retail, an eMarketer Podcast made possible by TikTok. This is the show where we talk about how retail collides with every part of our lives. I'm your host, Sara Lebow. Today's episode topic is real examples of AI in retail. Before we jump into those, let's meet today's guests. Joining me for this episode, we have Senior Analyst Blake Droesch. Hi, Blake.

Blake Droesch (00:52):

Hey, Sara. Good to be back.

Sara Lebow (00:53):

Good to have you. And back with us from the UK is Senior Analyst Carina Perkins. Hey Carina.

Carina Perkins (01:00):

Hi, Sara. Good to be on as always.

Sara Lebow (01:02):

Good to have you. Okay, let's get into AI. Back in December, US digital retailers said generative AI would affect their business this year more than any other trend. That's according to data from Bolt Inc. But a lot of the conversations we've had around generative AI over the past two years, even on this podcast, have been very theoretical, abstract, predictive. But a lot of retailers are using generative AI now. It's here. I want to get into these real-world use cases. Both of you recently wrote reports on this that are available on emarketer.com. How did you find that retailers are using generative AI?

Carina Perkins (01:41):

So retailers have really been experimenting with AI across a range of use cases for different things in retail. So one thing has been personalization, so really trying to personalize the shopper experience. Another thing has been improving search and the overall shopper journey, and then they've also been using it for customer service and to automate some of those customer service interactions and also help their teams out in-store and online.

Sara Lebow (02:06):

So let's break down some examples of those. Carina, you wrote your report specifically about generative AI for personalized marketing. Can you tell us about a company that is actually doing that?

Carina Perkins (02:17):

Yeah, so Coca-Cola is not a retailer, but it's a brand that's done quite a lot in this space. It was one of the first companies to engage with the alliance between OpenAI and Bain & Company, which was formed after the launch of ChatGPT to really accelerate generative AI adoption, and it's done this in several ways. So it created the Real Magic platform, which enabled people to create their own artwork and creatives using assets from its archive, but it also has used AI to create hundreds of variations of different marketing creatives and then A/B test them. It's even developed a new flavor, although that's arguably not hugely personalized, but it's done a lot of personalized marketing. And I think personalization and marketing, it really plays into generative AI's strengths. It really, really excels at content creation, at creating lots of content, be that text, image, video, really quickly and it can automate a lot of the tasks that marketing professionals are having to do manually at the moment, and it really frees them up for more creative and strategic thinking.

Sara Lebow (03:21):

That's not far from stuff Coca-Cola has done in the past. I mean, they've had the personalized cans with people's names on them for years and years, but are people finding this creepy? I think I'd be wigged out if I got a Coca-Cola ad that was like, "Hey Sara, here's a polar bear."

Carina Perkins (03:39):

Yeah, I mean, I don't know. It's difficult to tell. There haven't been many consumer studies in terms of the personalized marketing and how they feel about it, or not that I've seen.

(03:48):

But I think another interesting thing that companies have started doing, I think PMA just announced that it was doing, is creating personalized copy on landing pages. So if you go and hit a landing page, it will tailor the copy to you and your particular interests. And I think that can be potentially really powerful. People probably don't even realize that it's been personalized to them. It's not necessarily really obvious to them, but it will speak to the things that they're interested in and things that are going to engage them.

Sara Lebow (04:16):

So if I bought shoes for racquetball in the past, I might get a copy about racquetball?

Carina Perkins (04:22):

Yeah, exactly.

Sara Lebow (04:23):

That makes sense. Okay, let's move a little further down the funnel into conversational search. Blake, I know you've been researching that. How are companies using generative AI in conversational search?

Blake Droesch (04:35):

Yeah, so I think so far we've seen a lot of the major retailers roll out conversational search on their platforms, whether integrated with chatbots or not. Target, Amazon, Walmart, smaller retailers like IKEA. Really, I mean this is the forefront, at least in my opinion, of consumer-facing generative AI tools within e-commerce. I mean, really what it does is it puts the technology front and center and is asking shoppers to interact with the tool in order to really guide their shopping experience from the get-go.

(05:17):

And I think where we are is yes, there is implementation, but as Carina alluded to, we don't really have a great sense of how effective it really is at this point. I mean, we've got some clips from what retailers are self-reporting on how it's performing, but there's really not a lot out there in terms of how it's changing the shopping experience and how it's actually impacting conversion. But the idea is basically that conversational search is going to improve just the overall dynamics of the search engine on e-commerce websites with the large language model that's replacing standard tag or keyword-based search, but also that it could open up a new type of thematic search where instead of the classic example of you're planning, let's say it's October, so a Halloween party. Instead of searching for decorations and cutlery and candy, you could just say, "I'm planning a Halloween party," which would open it up to all of the different products that one might need. But there's still a lot to be seen as to how effective it's actually going to be in terms of consumers using it.

Sara Lebow (06:34):

Also, even if that tech is effective, people actually have to change how they search.

Blake Droesch (06:35):

Which is hard.

Sara Lebow (06:40):

Some of this is retailers catching up to how folks search on Google. We search with questions a lot on Google rather than strictly Boolean search terms, and a lot of retailer search interfaces aren't as good at those natural language questions. But getting into having a conversation with a search engine, that feels a little more foreign to me as a consumer.

Carina Perkins (07:05):

Yeah, and I think that's true. And I think that we do tend to see consumer adoption of new technologies is generally quite slow. If you look at augmented and virtual reality technology, the actual adoption rates are still fairly low, especially among non-Gen Z consumers. So I think there is going to be a bit of a lag between retailers rolling this out and consumers adopting it. And I think one thing that's really going to prevent adoption is if there are additional frictions added in because of the new technology. And if the conversation is frustrating and doesn't get them to where they want to be, I think that they'd probably revert to just wanting to search as they always have.

Sara Lebow (07:44):

Yeah. It's interesting you bring up AR and VR because I was talking to some folks at Snapchat about this last week about how common AR is on Snapchat. And I feel like a lot of people use AR without realizing they are AR users. They use filters. They don't realize that that is using AR. That seems like the best-case scenario for conversational search. You start using it without realizing that you're a generative AI user and suddenly, this is worked into how you shop.

Blake Droesch (08:15):

I think there's specific use case for this, and I think 90% of the time when we are shopping online, we know exactly what we want and we're just trying to get there as fast as possible. And LLM-based search can definitely improve that experience, but it's really... The question is, what is that other 10% when you're really not sure what you're looking for? You're not looking for a specific product, but maybe you're looking for a new pair of running sneakers and you're going to look to conversational search in order to hone in on exactly what you're going to end up buying. How likely are consumers, shoppers willing to take that route, really relying solely on the silo of conversational search rather than explore all of the many other tools that they have at their disposal to research different products?

Carina Perkins (09:15):

I think it is something, though, that is going to change slowly. And I think as we see more integration of AI and conversational AI into smartphones, for example, people are going to get a lot more used to integrating AI or using AI for everyday tasks. So that might give it a bit of a push, but I still don't think we're going to see really rapid consumer adoption of these things. And I don't know if any of you have tested out some of the AI shopping assistants that are available, but some of the user experiences are still quite frustrating. And I think the risk of companies launching too early is that they're just going to put consumers off the technology.

Sara Lebow (09:55):

Carina, you wrote about virtual assistants. How is that different from conversational search?

Carina Perkins (10:01):

Well, I mean it's not really. There's a lot of overlap in a lot of these use cases.

Sara Lebow (10:06):

Mm-hmm.

Carina Perkins (10:06):

Conversational search is specifically potentially talking about that could be integrated into chatbot, as Blake said, or it could just be replacing the standard search bar on a retail website, whereas a virtual shopping assistant would be more of the chatbot format. And potentially, it could also integrate customer service into what it does. So you could get help from this kind of virtual shopping assistant, right, from asking product discovery through to completing the purchase and helping with delivery and orders and things like that.

(10:39):

So you've also got, then, personalized product recommendations, which are a use case on their own. People are already getting AI product recommendations. Generative AI can improve the recommendations by real-time data and contextual factors. So a lot of these use cases do overlap, and I think we'll see retailers using a variety of them.

(11:00):

Some of the AI-powered shopping assistant examples is Amazon's Rufus, which can answer product questions, provide comparisons, and make recommendations. And then Zalando, which is a fashion retailer, has launched a ChatGPT-powered fashion assistant where customers can go and ask, say, "I'm going to a wedding in Tuscany in May. Can you suggest an outfit?" So there's an overlap there with search. It's quite hard to distinguish all of the use cases, really. There's a lot of overlap.

Sara Lebow (11:29):

So Rufus is homegrown from Amazon, but it sounds like Zalando's interface is through ChatGPT. Are these search interfaces, virtual assistants available to smaller brands that might be considering them, or is the activation cost too high?

Carina Perkins (11:48):

Yeah, I mean, setting up a custom large language model is expensive and time-consuming. LLMs are really complex and they have quite high computational requirements and challenges around processing power. So it's probably out of the realms of possibility for smaller retailers to build a custom solution, but we are seeing a lot of out-the-box solutions being launched now.

(12:11):

So one example is Dynamic Yield by Mastercard. It's launched Shopping Muse, which is a gen AI tool that is basically a virtual assistant that people can implement on their own websites. And Shopify has a ChatGPT-based shopping assistant. Google Cloud has a virtual agent. And I think there are similar things in search. Aren't there, Blake?

Blake Droesch (12:33):

Yeah, there is a, in the OpenAI GPT store, a out-of-the-box conversational search tools that smaller retailers can access. It runs the gamut, right, because Amazon, Target, they've decided to basically build their own platforms from scratch. Walmart is an example of a obviously massive retailer with deep pockets to invest in this technology. They've been investing in AI for... They're really early adopters, but their generative AI, their LLM platform was still built with some collaboration from OpenAI as well.

(13:13):

So I think there's really... It is the wild west in terms of what the best practices are for retailers of different sizes to go about implementing this type of stuff. The thing that I'm really going to be watching closely is there are a lot of these out-of-the-box solutions now. The question is, how well do they work, right? And I think in order for this type of technology to scale, that out-of-the-box solution's really going to have to be seamless.

Sara Lebow (13:42):

This might be a controversial take, but I think there an advantage to being a second mover here and waiting for Walmart, Amazon to play the field, see how consumer behaviors develop, see how the tech develops, and then move into that space. I know that a lot of people and people we work with are very bullish on moving faster with AI, though, so might not be the best advice.

Blake Droesch (14:07):

Yeah. I mean, I put in my report basically the same thing. I mean, if you're rolling out a tool that's really prominent, particularly in the search bar, you really run the risk of upsetting a lot of people with a really bad experience that it also is taking away from something that they're really used to, right? So being able-

Sara Lebow (14:08):

Yeah.

Blake Droesch (14:30):

... to use a search engine. I think that there really is.. It is such an arms race right now because everyone does want that first mover advantage. And yeah, it's true that if there's a shopping platform that can really get consumers engaging with generative AI tools, that's going to be a really big advantage for them. But it's also a huge risk. I mean, if you stumble out of the gate, then not only are people going to be upset with you as a brand, but they're going to be turned off in the future when it comes to trying some of these other tools that retailers are really pushing.

Carina Perkins (15:08):

Yeah, and I think that there's some things that people need to think about when they're choosing whether not to implement and what to prioritize. And I think like Blake said, if you're replacing one consumer experience with the other, then you need to be really careful. Because if you don't get it right, you're just going to create a load of friction and frustration. For example, Zalando's ChatGPT tool, that's a nice little extra, you know? So if it doesn't-

Sara Lebow (15:32):

Mm-hmm.

Carina Perkins (15:32):

... give you exactly what you were looking for from a fashion point of view, you say, "Oh, well. I'll just revert to the normal search." It doesn't get in the way of any normal processes that I would've followed, whereas if you are replacing your customer service with a ChatGPT-powered bot or a generative AI-powered chatbot and it gets something wildly wrong and gives your customer a wrong answer or gives a toxic answer or any of these risks that come inherent with the technology, you potentially then have a really damaging situation on your hand. So I think there are safer use cases where retailers can test it out and test the water, and then there are riskier applications.

(16:12):

And, I mean, arguably anything customer-facing is riskier. We've also seen retailers developing employee assistance chatbots. So the employees in-store or if they're customer service agents, can get a bit of help and pull up customers' backgrounds and things like that. But even those can be a bit of a risk because for the employee, if you're adding a load of frustration into their daily task, if you're not considering the workflows, if they're not properly trained on the tool, that is also potentially a kind of operational risk. So there's a lot to think about when you are implementing it.

Sara Lebow (16:48):

Before we get a little more into the nitty-gritty of when to use and not use generative AI, I want to talk about one more area of gen AI implementation we haven't gotten to, which is merging text and image-based search. That's easier to say than I just made it sound. But Blake, can you talk a bit about retailers that are using that?

Blake Droesch (17:08):

Yeah, this is one of those bleeding edge use cases that isn't really widely available, but the idea behind using generative AI to merge text and image-based search basically is taking the fact that large language models can use different types of media, not just text, to basically converse with the user. So the idea behind image-based search, obviously something that's become very popular over the last couple of years, is this idea that you could basically use text to augment an image-based search to basically have a multimedia type of search query.

(17:52):

So for example, if you saw somebody out on the street wearing a jacket that you really liked, but it was a red jacket and you wanted it in blue, you could upload that picture of the red jacket and say, "Find me a jacket like this but in blue." And this is a technology that Google is working on right now, and it's in an experimental phase and it hasn't really been rolled out widespread yet. But there's another AI startup called Daydream that is also in the early stages of rolling this out with a bunch of retail partners, including Alo Yoga and several other fashion brands.

Sara Lebow (18:34):

This is something that feels to me like more of an asset to a Google than an Alo Yoga because Google has every product in there. So if I'm like, "I want this jacket in red," Google can give me something. Alo Yoga, if I'm like, "I want this jacket in red," well, if Alo Yoga doesn't make that jacket in red, then that's the end of that.

Blake Droesch (18:55):

Yeah, I think it's definitely... I think a lot of gen AI search is meant to optimize the long tail, which is really useful for Google, but also eBay, ThredUp, a lot of marketplaces that have a really, really large consideration set of products. I think you're right. An Alo Yoga might not necessarily, something with smaller inventory, might not necessarily be great for hammering down those exact sort of product specifications. For a retailer like Alo, I think it's really more about finding products with certain specifications that are going to match the needs of what the person's searching for like, I don't know, for example, something like running shoes that are good for people with bad knees, but not necessarily nailing down a specific, niche product.

Sara Lebow (19:53):

The last thing I wanted to talk about is just when should you not use generative AI? A lot of people will be like, "This is going everywhere. This is good for everything." It's not. What are the places that it shouldn't be used?

Carina Perkins (20:04):

Yeah, so I think we've seen everyone jump on generative AI because you could potentially use it for endless, endless use cases, but there's some things it's really not that great at. So it still has a lot of issues around accuracy, around the explainability of some of the answers it comes up with, around bias and things like that. So it's actually not the best at demand forecasting, at inventory optimization, certainly not at high-level decision-making, and you wouldn't really want to give it any autonomy at this stage in decision-making. There are other forms of AI which are better suited for that. So predictive AI, which uses algorithms and statistical models to predict future trends, that's much better at demand forecasting.

(20:47):

But what we are starting to see, which I think quite interesting, is companies combining the two technologies. So you use the predictive AI for the forecast, and then you benefit from the conversational interface of generative AI to explore that data. And it's also really good at servicing knowledge, generative AI. So I think we're going to see going forward a bit more, rather than just trying to get generative AI to do everything, people trying to combine the two technologies. And we've already seen software solution providers offering products that do combine the two.

Sara Lebow (21:20):

Yeah, that makes sense. Okay, well, that is all we have time for today. So thank you both for being here. Thank you, Blake.

Blake Droesch (21:27):

Yeah, thank you.

Sara Lebow (21:28):

Thank you, Carina.

Carina Perkins (21:29):

Thanks, Sara.

Sara Lebow (21:30):

Thanks to our listeners and to Victoria who edits the podcast so we can generate an episode for you each week. We'll be back next Wednesday with another episode of Reimagining Retail, an eMarketer podcast made possible by TikTok. And tomorrow, join Marcus for another episode of the Behind the Numbers Daily.

First Published on Oct 2, 2024