Retailers want AI to tell them where to put new locations. The tech isn’t there yet
Artificial intelligence is widely used to forecast sales trends or foot traffic, or to help retailers understand how to merchandise their store, or what customers are buying or looking at online.
There is, however, one area in which retailers have been underwhelmed by AI as a predictive tool. It’s real estate, executives at major brokerage firms that represent some of the nation’s biggest retailers told Modern Retail. They said their retailer clients are looking to AI to predict which locations will be more profitable, but that the technology and underlying datasets aren’t yet capable of doing so.
“AI is really good at parsing large sets of data, helping us delineate some insights from it, so we can use AI type models in retail to help with how we operate the business,” said Paul Sill, head of the Visionary Insights Group at JLL. “But in terms of how we pick sites and the real estate questions of, ‘Hey, is this corner going to be better for me than that corner?’ AI isn’t really moving the needle as much as people want to believe it is or think it can.”
Sill leads a team of data scientists and software engineers who help retailers use their in-house data in conjunction with JLL’s market and consumer data to predict their future performance. He has been in retail analytics, site selection and sales forecasting for about 30 years and has worked for retailers, as well, including Blockbuster and Sears.
“Everyone’s in a rush, and has been in a rush, to see AI have some dramatic impact on the ability for retailers to more accurately predict success, and predict sales and performance of their stores,” Sill said.
Scott Schnuckel, managing director of Americas retail for commercial real estate brokerage CBRE and former longtime head of real estate for Kohl’s, said AI has helped expand retailers’ ability to access more data and from a mobile device while scoping out sites. With new technology, companies can factor more data sources — even Facebook impression counts, for example — into real estate decisions.
“Ten to 20 years ago, we would go find our 10-20 key variables across the demographics, we’d apply some methodology to it, and then we’d use that to help us assess sites,” Schnuckel said. “Now you’re talking about hundreds, maybe even thousands of variables — some that are even being picked up without you, the user, even thinking that may be a variable that’s important.”
But Sill said, because AI lives off of large data sets, retailers often don’t have enough data from different scenarios to substantiate whether one decision would be better than another. The success of a given location is driven by many different aspects, from the site itself to the competitive dynamic in the marketplace, and even the way stoplights are set up or how traffic patterns flow, Sill added.
“If I’m a 500-unit coffee shop, 500 observations is not enough data for AI to do anything intelligently. AI needs millions of data points,” Sill added. “There are so many things that are part of that overall success matrix of what makes a particular site successful. AI alone and where its limitations are mean we have to go way beyond just thinking AI can solve the problem for us. There are just other ways we can model data to be more successful with predicting site performance.”
For smaller chains that don’t have as much data to work with, Sill and his team rely on other machine learning techniques — largely, multivariate regression modeling — rather than AI, which uses decision tree and neural net models. Through the regression modeling, Sill said, retailers can drop a pin on a map and predict with some known degree of accuracy that a site will do $5 million in sales versus $4 million. It’s still not an exact science.
“It’s about taking the next 100 potential places I might build a store and using a model to tell me that this top 20% has a much lower risk profile for me and has a higher rate of return,” Sill said of the regression modeling. “It won’t be perfect, but again, if you approach it with that technique and that methodology, you are minimizing your risk exposure and maximizing your return on investment.”
AI is also limited, in that there are so many other human elements that lack patterns that it may never be able to pick up on. “You can have a great site and great market with a crappy manager who has staff turnover,” Sill said. “Some of the work we do for clients is about identifying where those gaps exist.”
Even if it isn’t yet fully capable of predicting the success of a store at a given site, AI-based tools can be helpful in helping retailers find trends in vast swathes of data — especially when they have their own robust set of customer transaction data. Retailers like Walmart and Kroger have been building out data science divisions over the last few years that help to inform their own executives and merchants, as well as suppliers.
“We’ve had some success mining millions of transactions of e-commerce data, millions of transactions of customer household data to figure out the patterns in that data that differentiate where they’re more successful or less successful,” Sill said. “That can be used to translate into some of the site selection modeling.”
One popular AI-powered tool often used by retailers and brokers is Placer.ai, which partners with apps to obtain location data that customers have opted into sharing, minus any personally identifying information. With these data points, the platform can track how many people have visited a grocery store or gym, for example.
Caroline Wu, director of research at Placer.ai, said AI has enabled Placer’s engineering team to create a platform that synthesizes a massive set of location data. As a result, retailers or brokerages can compare foot traffic now to five years ago, look at competitors’ traffic or determine the busiest days of the year, among other use cases. This can help with site selection, as well as planning promotions, marketing or events.
The platform, still, is dependent on retailers and their brokers knowing what they want and understanding their customer base. They have to manually input parameters into the platform, so it’s a more manual process than, say, asking a chatbot where to put a location to be as profitable as possible.
“The retailer does have to have some idea of what they’re hoping to accomplish,” Wu said. “Are you hoping to grow in particular markets? … If you’re a gym, you’re going to want a much larger footprint than if you’re a smaller boutique apparel brand. Those are some of the parameters that you as a retailer would need to put in.”
Wu doesn’t imagine Placer will replace the thinking or expertise of real estate brokers anytime soon. The platform’s long-term goal is to continually refine retailers’ understanding of the drivers of successful retail or real estate decisions, Wu said. She hopes to continue shortening the time it takes to make a real estate decision by synthesizing Placer’s cell phone data with other data, such as demographic information and income levels. Placer also has an API so retailers can combine its data with their own proprietary numbers.
Sill said cell phone data can give JLL and its clients answers to specific questions, like how far people are traveling to visit the center that you’re looking to be in or where that future location’s customers would likely be shopping today. But it’s detached from the transaction itself — so it lacks any context on what people who visit a location are actually buying — and is limited to customers who have opted in to whichever apps such platforms have partnered with.
“It doesn’t help us predict sales by itself,” Sill said. “It just becomes part of our overall machine learning model that explains some small portion of your sales variance.”
Schnuckel similarly said AI vendors have yet to successfully bridge together the many different variables to successfully predict whether a location would be profitable.
“Getting to an all-in-one data lake to be able to operate off them — I’m not aware of anyone who’s doing that real well yet,” he said.