→ How the model works

The thesis

Site selection was never an AI problem

By The Big Data Company · Reviewed July 2026 · 5 min read

The hard part of a site decision was never the algorithm. It was getting honest data about who is really within reach, and a method you can argue with.

Every location-intelligence pitch now opens with the same word. AI. More of it, applied to more data, producing a more confident number. It is a comfortable story to sell and a comfortable one to buy, and for forecasting the revenue of a store that does not exist yet, it is mostly the wrong story. The hard part of a site decision was never the algorithm. It is getting honest data about who is really within reach of a site, weighing it the way people actually travel, and comparing the candidate to the stores you already run.

A black box that cannot show its work cannot be argued with, corrected, or trusted with a lease. That is not a small inconvenience. A store decision is a multi-year capital commitment defended in a room full of people who will ask why. A number with no reasoning attached is not an answer to that question. It is a request to stop asking it.

So IRIS uses machine learning where it earns its place, and keeps the model one you can open: interpretable, grounded in your own estate, and open to the one adjustment that is ever valid, when it is missing something you know. The reason for that posture is not taste. It is arithmetic.

The arithmetic nobody in the pitch mentions

A model asked to learn geography from scratch, who can reach a site, how demand splits across competitors, out of raw signals alone, needs an enormous number of examples to do it without simply memorising the ones it saw. Several independent lines of reasoning all land in the same place. The classical rule for not overfitting even a plain linear model is on the order of ten to twenty observations per predictor, and IRIS reads more than a hundred local signals per location. A model that also learns the interactions and representations among those signals, which is what "end-to-end" means, has a vastly larger space of possibilities to pin down, and wants orders of magnitude more.

Put a number on it and the gap is stark. A from-scratch model of this kind sits comfortably in the regime of tens of thousands to millions of comparable labelled stores. A brick-and-mortar operator does not have that. One store yields roughly one independent label; extra years of the same store are correlated, not fresh draws; and pooling stores across formats, countries, and eras breaks the comparability the model needs.

Read the two side by side. A typical chain sits in the hundreds, two to three orders of magnitude below what a from-scratch model needs. Even the very largest global operator only appears to reach the bottom of the band, and only by counting restaurants across dozens of countries and formats that no single-market model could honestly pool together. On the data a real expansion team has, a pure end-to-end model does not learn the geography. It fits the noise.

The alternative is not less rigour. It is the right rigour

The way out of a thin-data problem is not a bigger model. It is a better prior. More than a century of work in economic geography has already established how trade behaves: a location earns more when more people can reach it easily and are inclined to spend once they do. That regularity is one of the most tested in the field. Build it into the model before you look at the data, and you no longer have to spend your few hundred stores rediscovering that distance matters. You spend them sharpening a structure that already knows it.

This is where machine learning belongs, and it is real work, not decoration. It reads the hundred-plus signals per location, and it learns the embeddings that place similar sites near each other so the model can transfer what it learned at one location to a comparable one. What it does not do is replace the forecast with a number no one can explain. The learning is bounded inside a structure you can read.

What you are left holding

The payoff of that choice is not abstract. Because every part of the formula maps to something real, demand, attractiveness, distance, competition, a forecast can be read, questioned, and checked rather than merely believed. Because the model is defined over competing destinations, you can add a competitor or open one of your own stores and recompute who captures what; cannibalisation and scenario planning are the same model run again, not a feature bolted on. And because the structure is stable, small changes in the inputs produce small changes in the forecast, not the wild swings a data-hungry model shows when you feed it a few hundred rows.

None of this is a reason to distrust AI. It is a reason to be specific about where it helps. The rest of this series walks the method in order: the gravity model behind the number, what actually goes into a forecast, how IRIS validates it out of sample, and, honestly, where the model is weakest.