In short: Greenfield analysis starts by deleting everything you own, so the question becomes where warehouses would go given a demand map and a freight cost structure rather than which building anyone is willing to defend. For a single site the answer is the demand-weighted centre of gravity, and for several it is a partition of demand points with a site at the weighted centre of each group. The site count is a commercial quantity rather than a geometric one, since the algorithm answers for any number you ask for and has no view on which is right. The gap between your current network at its actual site count and the greenfield answer at the same count is the price of your history.
The network review starts with the wrong question almost every time. Somebody opens with which of our nine sites should we close, and the next ninety minutes are spent defending buildings. The site director with the best relationships wins, the site with the newest automation wins by default because writing it off is unthinkable, and the site everyone privately knows is in the wrong place survives because nobody wants to be the person who closed it. Whatever comes out of that meeting is bounded by decisions somebody made in 1998.
Greenfield analysis changes the opening question. Delete the estate. You own no buildings, no leases, no conveyors and no people. You have a demand map and a freight cost structure. Where would you put warehouses, and how many.
The answer works as a benchmark rather than as a plan. It is the only way to put a number on what your history costs you, instead of arguing about it in anecdotes.
Start with the point that minimises weighted distance
For a single site the question has a clean formulation. Every demand point has a location and a weight, and you want the point minimising the sum of weight times distance to every demand point. That is the demand-weighted centre of gravity, and it is worth being precise about it because the version most people run is a different calculation than the one they think they are running.
The true minimiser of weighted straight-line distance is the Weber point, also called the geometric median, and there is no closed form for it. You get it by iteration, usually with the algorithm Weiszfeld published in 1937, which repeatedly recomputes a weighted average where each weight is divided by the current distance to that point, and converges quickly. The formula that appears in every network spreadsheet, where you take the volume-weighted average of the latitude column and the volume-weighted average of the longitude column, minimises weighted squared distance rather than weighted distance. That is the centroid rather than the geometric median, and squaring pulls the answer harder toward distant heavy points. On a national demand map with one remote industrial cluster the two answers can sit a long way apart, and the centroid is the one that drifts toward the outlier.
Three other details change the answer more than people expect. Averaging latitude and longitude is acceptable over a couple of hundred kilometres and wrong across a continent, so project properly or work on the sphere. Straight-line distance is not road distance, and the ratio between them runs somewhere around 1.2 to 1.4 in ordinary geography and much worse around mountains, estuaries and borders, so a candidate point that looks central can be an hour further from half your demand than the map suggests. And the weight should be freight cost per unit of distance rather than raw volume, because a tonne moving by parcel and a tonne moving by full truckload cost very different amounts to move a kilometre. Weighting by volume alone sites the network for the wrong flow.
From one site to k of them
One site is rarely the question. The multi-site version is: partition the demand points into k groups, place a site at the weighted centre of each group, assign each demand point to its nearest site, and minimise the total weighted distance across the whole map.
That objective is exactly what k-means minimises, with demand as the weight. The algorithm alternates between assigning points to the nearest centre and recomputing each centre from the points assigned to it, which is Lloyd's procedure, named k-means by MacQueen in 1967. It is fast enough to run at postcode grain on a laptop, which is the property that makes it useful for the sweep described below.
Two caveats belong in the same breath. K-means minimises squared distance, so it inherits the centroid behaviour described above and is pulled by outlying demand. If that bothers you, and it should if your demand map has a long geographic tail, run the k-medians variant that recomputes each centre with Weiszfeld instead of by averaging. And the procedure converges to a local minimum that depends on where the initial centres were placed, so a bad start produces a bad answer with no warning attached. Seed it deliberately with the k-means++ scheme Arthur and Vassilvitskii published in 2007, which spreads initial centres in proportion to the squared distance from the ones already chosen, then run it from a number of seeds and keep the best. Record the seed alongside the result, because finance will ask you to reproduce this number in three weeks and a different answer will end the conversation about whether the analysis is real.
What the k means commercially
The algorithm will happily give you an answer for any k you ask for, and it has no view about which k is right. That is appropriate, because k is a commercial quantity rather than a geometric one. It is the number of buildings you are willing to pay fixed cost on, staff, fit out and manage, and the geography follows from it.
So the useful discipline is to stop looking for the k the model recommends and start reading the curve. Run the clustering at every k from one to about twelve. Plot total weighted distance, converted to freight cost, against k. It falls monotonically with diminishing returns, steeply from one to four and slowly after six for most national networks. Then draw the annual fixed cost of a site as a straight line in k and add the two. The sum is U-shaped, and the bottom of that U is your first estimate of how many sites the demand map supports.
Read that bottom carefully. It is usually flat across two or three values of k, which means the honest output is a range and deserves to be presented as one. The elbow people expect to see in these curves is almost always softer than the slide that eventually presents it, and a k chosen from the third decimal of a flat region is a decision dressed up as a finding.
The greenfield number matters more than the greenfield map
Here is the part worth more than the pretty map with dots on it. Take your current network, hold the site count fixed at what you actually operate, and compute the same weighted distance cost. Then take the greenfield answer at the same k. The difference between the two is the cost of your history, in currency, per year.
That number decides whether a network programme is worth running before anybody spends a month on it. A gap of two or three percent means your predecessors chose well, the map has not moved much, and the file should close. A gap in the mid teens means there is a real case, and it should go into the constrained facility location model where capacity, real candidate buildings, service constraints and the actual cost of opening and closing get considered properly. That is a separate exercise with its own method and its own literature, and the greenfield step exists to tell you whether it is worth starting.
Presenting it as a benchmark also protects the analysis from the most common objection. When you show a greenfield map to an operations audience, the first response is that you cannot put a site there, and they are frequently right. That objection lands harmlessly against a benchmark. It is fatal against a proposal.
What moves a greenfield point in practice
The geometric answer is a search centre. Draw a radius around it, ask a property adviser what exists inside that radius, and the buildings that come back become the candidate list for the constrained model. Several things routinely move the real answer away from the geometric one.
Labour availability and its real price. A regional distribution centre needs several hundred people at peak. Check the catchment population inside a realistic commuting time, the unemployment rate, and how many competing operations are already hiring from it. A perfect geometric point in a market with three large fulfilment buildings already in it is not available at any wage you would agree to pay, and the cost shows up as agency premium and turnover rather than as a higher headline rate.
Land, buildings and power. Big-box industrial supply is lumpy and slow. The right site may be thirty kilometres from the geometric point because that is where a unit of the required size with the required power connection and planning consent exists. Power in particular has become a live constraint for automated buildings in several markets.
Transport infrastructure and driver hours. Motorway junction access, rail connection, and proximity to a port or air hub change the effective cost of a location well beyond what a distance matrix captures. Driver hours regulations make delivery reach a step function instead of a smooth one, so a location that puts a whole region inside a single legal driving day is worth considerably more than its kilometres imply, and a location twenty minutes the wrong side of that boundary is worth much less.
Tax and incentive regimes. Enterprise zones, property tax abatement, regional development grants, customs treatment such as a free zone or bonded warehousing. These can be worth several years of the freight difference between two candidate points, which is the argument for having someone from tax in the room early rather than at the approval stage.
What you already run nearby. Shared management, an existing labour relationship and a known local authority are real advantages, and they are the one legacy factor worth admitting into a greenfield discussion, clearly labelled as the exception it is.
Where this stops
Greenfield analysis ignores everything you own. That is the entire point of it, and it is also the reason it can never be the answer by itself.
It has no view on the freehold you own outright with no exit cost, the lease with seven years to run and a dilapidations clause, the automation you paid for four years ago, or the workforce you would have to recruit and train again somewhere else. A greenfield answer implemented literally means writing off assets whose sunk cost is irrelevant to the model and extremely relevant to the year in which you take the charge. It also has nothing to say about capacity, because it puts a point on a map without asking whether a building of the necessary size can exist there or whether one site can physically handle the volume assigned to it.
It is a static picture as well. You fed it one year of demand, and the demand map moves, sometimes quickly when a large customer changes its own receiving structure. Running the analysis on last year and on the year before that is a cheap way to see whether the centre of gravity is drifting and in which direction, and drift matters more for a twenty-year property decision than the current position does.
The last honest limitation is that distance is a proxy for cost. Weighted distance is a good proxy on a single mode with roughly uniform rates, and a poor one where your rate structure has zone breaks, fuel surcharges, minimum charges on small parcels and different carriers by region. Where the rate card is complicated, the greenfield curve tells you the shape and the constrained model with real rates tells you the level.
This week, pull last year's outbound shipment lines with destination postcode, weight and freight paid, aggregate them to a demand map, run the weighted centre of gravity on freight cost rather than volume, and measure how far your largest site sits from it.