Journal / Kirill Naumov
AI for Retail Site Selection: Screen Better, Then Walk the Street
How retailers can use machine learning to shortlist potential store locations—without mistaking a model’s ranking for a lease decision.
By Kirill Naumov ·
A location model can help a retailer sort a long list of possible sites. It cannot tell you what the morning queue feels like, whether the entrance is visible from the street, or whether a landlord’s promises survive a site visit. Use AI to decide where to look first—not where to sign.
What the research does—and does not—show
A 2024 study in the [Journal of Retailing and Consumer Services](https://doi.org/10.1016/j.jretconser.2023.103620) tested machine-learning approaches for early location screening using data from 743 jewellery stores in China. The authors combined public city, competitor and point-of-interest information with other market indicators to rank candidate areas. That is useful evidence for a screening workflow in that setting; it is not proof that the same model, variables or ranking will transfer to another category or country.
A separate [Yonghui Superstores case study](https://doi.org/10.1108/MSCRA-03-2019-0010) describes combining location information, population, competitors, transport and the retailer’s own store network for site selection. The lesson is not “let the algorithm pick the shop”. It is that a decision improves when external context is considered alongside the retailer’s own operating experience.
Give the model a shortlist job
Start with a defined decision: which areas deserve a closer look for a new convenience shop, beauty store or showroom? Build a candidate set, then use relevant, lawful data such as travel access, local demand indicators, nearby destinations, competition, existing store reach and the format’s own historical performance. Keep the data date and geographic scale visible. A city-level score should not be presented as a street-level truth.
Ask the model to rank or group candidates and show the factors that moved them up or down. Compare its list with a simple benchmark—current planning practice, a transparent scorecard or a small set of known locations. If the model cannot beat a clear baseline, it has not earned a place in the process.
Keep the investment decision human
A shortlist is an invitation to investigate. A local team should check pedestrian movement at different times, sightlines, access, nearby construction, unit condition, competitor changes, catchment boundaries and the practical cost of staffing and delivery. Finance and property teams still need to test rent, fit-out, lease terms and downside scenarios. Add what the data cannot see, then record why a candidate was advanced or rejected.
Be careful with proxy variables. Population, mobility or neighbourhood attributes can reflect access and opportunity, but they can also reproduce historical bias or hide who benefits from a location. Use only data that is appropriate for the decision, review for uneven effects and avoid individual-level location histories where aggregated measures will do. A model’s explanation is a debugging aid, not a guarantee of fairness.
Test before the next lease
Take a past cohort of sites and recreate the information that would have been available at the time. Compare the model’s ranking with later outcomes, while accounting for store format, opening date, local competition and major changes. Then shadow the next real expansion decision: record what the shortlist changed, what the site visit found and which unknowns still mattered. Do not judge the model only by the stores it helped open; include rejected candidates and locations where the model was confidently wrong.
A useful first pilot has one geography, one store format, a named property decision owner, a baseline method and a stop rule. Track research time, shortlist quality, field-validation findings and the gap between expected and observed performance. No single score can remove uncertainty from a lease. It can help a team spend its scarce attention on the most promising questions.
AI can make retail expansion research more systematic. The final call still belongs to people who understand the neighbourhood, the customer and the economics of the site. For other practical applications across retail, sales, staff learning and loyalty, see the [AI for Retail workshop](https://www.kirillnaumovai.com/workshops/ai-retail-november).