GIS & Spatial Analysis
Site Selection
Site decisions are among the most expensive and least reversible a distribution or retail business makes, and they are routinely taken on a combination of a radius map, a footfall estimate and whoever advocates most confidently. The lease runs a decade; the analysis took an afternoon.
A defensible answer needs three things the radius approach cannot give: a catchment bounded by how people actually travel, a demand model that allocates that population between you and the alternatives available to them, and an explicit estimate of how much of the new site's revenue is transferred from sites you already own. The third is the one most often skipped and the one that most often turns a good-looking site into a bad investment.
How the work runs
Four phases, and what comes out of each.
2–3 weeks
Build the demand surface
Before evaluating any site we build a picture of where demand physically is, population and its composition, existing customer geography from your own transaction data, competitor locations, and whatever proxies fit the category. Your own sales data geocoded to customer location is usually the single most informative input and usually the least used.
Outputs
- Demand surface at a workable spatial unit with sources documented
- Existing customers geocoded and profiled by origin
- Competitor and alternative-provider layer
2–3 weeks
Model real catchments
Catchments are built as travel-time isochrones over the actual road network, not radii, and calibrated against where your existing customers genuinely come from, which frequently reveals that the assumed catchment is both smaller and a different shape than believed. Where the category warrants it we fit a gravity model so that demand is allocated by relative attractiveness and distance rather than assigned wholesale to the nearest site.
Outputs
- Drive-time isochrones per site, calibrated against observed customer origins
- Gravity or share model allocating demand between competing locations
- Catchment overlap quantified across the existing network
2–3 weeks
Score candidates net of cannibalisation
Each candidate is scored on modelled revenue, then the transfer from existing sites is subtracted to give a net figure. This is the number that should drive the decision and it is routinely absent from site papers. It is also frequently the difference between a site that adds twenty per cent and one that adds four.
Outputs
- Candidate ranking on modelled gross and net revenue
- Cannibalisation estimate per candidate against the existing network
- Sensitivity of each ranking to the model's key assumptions
2 weeks
Hand over the model, not the answer
Site questions recur constantly and an analysis that has to be commissioned each time will be bypassed under time pressure. We hand over the model so your team can score a new candidate in an afternoon, along with the calibration protocol for keeping it honest as the network changes.
Outputs
- Working model your team can run on new candidates
- Documentation, training session and calibration protocol
- Backtest of the model against sites you have already opened
What good looks like
The ranges we design toward.
15–40%
Catchment error
Typical overstatement when a radius is used in place of a drive-time isochrone in a dense or barrier-heavy geography.
10–30%
Revenue transferred
Share of a new site's revenue commonly cannibalised from existing sites in an established network.
±10–20%
Model accuracy
Realistic first-year revenue prediction error for a calibrated gravity model. Anyone promising better is overfitting.
Deliverables
What you are left holding.
- 01Demand surface and geocoded customer base, retained and reusable
- 02Calibrated drive-time catchments for every existing and candidate site
- 03Candidate ranking on revenue net of cannibalisation, with sensitivities
- 04Backtest against sites you have already opened, so the model's track record is visible
- 05The working model, handed over with training rather than kept as a service
What an engagement looks like
A representative shape.
Illustrative. This is the shape of work in gis & spatial analysis, not an account of a specific client. We do not publish named results, and we would be wary of a firm that publishes them without the client in the room.
10 weeks
- 01Situation
A distribution business was assessing eight candidate locations using ten-kilometre radius catchments, and had recently opened two sites after which nearby existing sites declined without anyone having forecast it.
- 02What we did
Geocoded three years of transaction data to customer origin, built drive-time catchments calibrated against where customers demonstrably came from, and fitted a gravity model to allocate demand between competing sites.
- 03Where it landed
Modelled catchments were substantially smaller and differently shaped than the radii implied. Two candidates that had ranked highly on gross population fell sharply once cannibalisation was subtracted, and the ranking changed at the top.
Where this connects
Rarely a self-contained problem.
Questions
The ones we get asked.
Our addresses are messy. Can you still geocode our customers?
Usually, and it is one of the harder parts of the job in India. Unstructured addresses, inconsistent locality names and PIN codes that cover large areas all reduce match rates. We typically achieve seventy to ninety per cent geocoding at usable precision with a combination of matching strategies, then handle the remainder by aggregating to PIN or ward level rather than discarding it. We report the match rate and its spatial bias, because unmatched addresses are rarely randomly distributed.
How accurate can a revenue forecast for a new site really be?
For a calibrated model in a category you have many existing sites in, plus or minus ten to twenty per cent on first-year revenue is realistic. For a new format or a new region with nothing to calibrate against, it is considerably worse and we say so. The value is less in the absolute number than in the ranking. The model is far better at telling you which of eight candidates is best than at telling you exactly what the best one will earn.
Do we need to buy footfall or mobility data?
Sometimes, rarely at the outset. Your own transaction data geocoded to customer origin is usually more informative than purchased mobility data and you already own it. Commercial footfall data earns its cost in high-street retail where passing trade dominates; for destination formats and distribution it usually does not.
Can this handle dark stores and quick commerce?
Yes, and the modelling is if anything cleaner, because delivery radius is a hard operational constraint rather than a behavioural estimate. The interesting questions shift to coverage overlap, order density within the service polygon, and how many sites are needed to cover a city at a promised delivery time, which is a facility location problem and connects directly to network design.
Insights
Related reading.
Why a radius is not a catchment
Drawing a ten-kilometre circle takes ten seconds and is wrong in a specific, predictable direction. The lease it justifies runs a decade.
Cost-to-serve: the analysis most companies skip
Most businesses know their gross margin by customer. Rather fewer know which customers are profitable, and the two lists differ more than anyone expects.
Closely related
More on GIS & Spatial Analysis →Route & Territory Design
Balanced territories and efficient routes built on real travel times, so the workload is fair and the vehicles are full.
GIS & Spatial AnalysisDecision Maps
Maps that trigger a decision, in the browser, for the people doing the work, not another viewer nobody opens.
Start a conversation
Talk to us about site selection.
Most useful first conversations are thirty minutes and involve no proposal. Describe the problem you are looking at and we will tell you honestly whether it is one we can help with.
Or write directly to hello@shyenova.com
