Questions to ask any site-selection vendor

Every site-selection tool will show you a confident demo. The demo is not the thing you are buying. You are buying a number that a board will commit real capital against, and a method you will lean on for years. So before you sign, it is worth asking a few plain questions, of us and of anyone else you are considering. The answers separate a tool you can defend from one you simply have to trust.

Here is the short list we would want a buyer to bring to the table.

"Can you show me why a site scored what it did?"

Ask to see the reasoning behind a single number, not just the number. A good answer names the things that drove it: the residents and workers within reach, the passing trade, the nearby draws, and how each was weighed. You should be able to click a figure and trace it back to a source.

A weak answer is some version of "the model learned it". A score you cannot open is a score you cannot argue with, correct, or take to a sceptical board. If the reasoning is hidden, you are not being sold a method; you are being sold a verdict.

"Is it checked against my own stores?"

Any vendor can quote an accuracy figure before they have seen your data. It means very little, because it was not measured on your business. The question that matters is whether they will hold the model against the stores you already run, including ones it was never trained on, and show you how close it came.

A good answer welcomes that test and treats it as the starting point. A weak answer offers a single headline percentage and changes the subject when you ask which stores it was measured on.

"Does it give me a range, or a single number?"

A first-year revenue forecast is genuinely uncertain, and a tool that hides that is hiding the difficulty, not beating it. Ask for the range around the number, and ask how often reality is expected to land outside it. A site on the edge of your hurdle rate is a different decision from one comfortably above it, and only the range tells you which you are looking at.

A good answer hands you an interval as readily as a point estimate. A confident single figure, with no width, should make you more cautious, not less.

"Can a person overrule it, and does the model still run?"

You know things the data does not: a lease clause, a local quirk, a change coming to the street. A useful tool lets your own expert disagree for a real reason, change the number, and keep going, with the adjustment recorded. That is judgement kept in the loop, not designed out of it.

A good answer treats the human as the decision-maker and the model as the disciplined input. Be wary of anything that positions itself as the one making the call.

"Where does my data live, and do I keep the model and its outputs?"

Your store history is commercially sensitive, and the model built on it is an asset. Ask where the data is processed and stored, whether it trains only your model, and whether you can export the outputs and walk away with them. For a European operator, ask plainly whether it stays in the EU.

A good answer is specific and unbothered by the question. A weak one is vague about hosting, or quietly assumes you will be locked in.

The answers you want to hear

Notice that none of these questions is about who has the cleverest algorithm. They are about whether you can see the working, whether it was checked on your reality, whether it is honest about what it does not know, whether you stay in charge, and whether you keep what is yours. A tool that answers all five well is one you can defend in the room where the money gets committed.

That is the bar we hold ourselves to, and the bar we would want you to hold everyone to.