Client projectPrivate

Every competitor site in three markets, rescanned weekly and checked before it counts

5,700+ sites from eight public sources, rescanned every week, stopped at a gate when a list looks wrong, and turned into a map that plans where to build next.

Industry: Multi-market network operator
What it shows: Competitor intelligence, OSINT, Geospatial analysis, Expansion planning
Year: 2026
Status: Private

Illustration with a made-up region, sites and changes. The layers, change rules, scan checks and plan method are the project's real ones.

competitor and client sites in three markets, from eight public sources: 5,700+
conditions that stop a scan before a bad list can count: 19
same-place pairs right when checked by hand: 815 / 817
stores of 20 chains joined to the sites, checked against OpenStreetMap: 3,900

Figures come from the project’s own files and records.

What changes

The steps people did by hand, and the route the work takes now.

Management asks where competitors are opening

Before

Open every competitor’s map
Copy the sites into a spreadsheet
Guess what changed since last time
Eyeball the gaps

With AI

Eight public lists collected and checked every week
A planning map shows the gaps and where to build

A weekly change report and a build plan

The problem

The client competes on the ground. Its rivals open, close and move sites every week, and each publishes its network only on its own map, in its own format. Management wanted to know where competitors were opening, where customers walked past a rival’s site to reach none of the client’s, and where the next site would do the most good. Answering that by hand meant opening eight maps, copying thousands of rows and guessing what had changed since the last time.

What we built

A weekly census of every site in three markets, built only from public sources. Claude runs each scan from written instructions: collect the eight published lists as they are, turn them into one table, stop if anything looks wrong, compare with last week, and report what changed and where.

In the latest week the report found 45 changes: 30 new sites, 5 removed, 4 renamed, 3 with a new address, 2 moved and 1 with a refined location. The scan passed the gate with no stops and three warnings, each answered before the report went out.

Every scan that passes rebuilds a planning map. It joins the sites to an official 1 km population grid, area boundaries and 3,900 stores of 20 retail and fuel chains, and turns them into intelligence layers: who has no site within 1 km, where a competitor is within 1 km and the client is not, how each area scores, which chains host which operators, and where a new site would reach the most people the client does not reach yet.

Why the numbers can be trusted

A census is only useful if nobody has to wonder whether a change is real. A source that returns half its sites, a list cut off at a page limit or coordinates that come back swapped all stop the scan before it can count, and the stopped scan never becomes the baseline. History is never overwritten. A self-test runs a whole scan on planted data, including a broken source that must be stopped.

The map is held to the same standard. Before a new scan is applied, the rebuild must reproduce last week’s map exactly. Every published figure is then calculated three ways, in the page itself, independently from the scan and in the method text, and all three must agree. The rule for when two sites stand in the same place was checked by hand on 817 pairs and was right on 815.

The method also says what the data cannot do. Against about 2,400 sites mapped independently in OpenStreetMap, half of the operators’ own points sit within 13 m, three quarters within 30 m and nine in ten within 56 m. So a 25 m result is only indicative, while a 100 m result barely moves when the threshold does.

What it shows

This is open-source intelligence that management can act on: a standing weekly report instead of a one-off study, and a planning tool that ranks sites by the people they would reach rather than by a score nobody can explain. The same approach fits any business that competes through physical locations, from retail and banking to healthcare and charging. Below, explore a made-up region with the real layers, rules and checks.

Interactive · 8 journeys

Interactive system map

From eight public lists to a map that plans the next site

24 parts and 8 journeys: the weekly scan, the gate that stops a bad one, the person who checks every removal, and the map with its layers. Generic part names; the client stays anonymous.

Play 8 journeys, or click any of its 24 parts.

sites in three markets, from eight public sources: 5,700+
conditions that stop a scan: 19
same-place pairs right when checked by hand: 815/817

Explore the map

InteractiveSwitch the map's layers, filter this week's changes, size an expansion plan, or break a scan.

MapWhat changedWhere to buildThe scan gate

A made-up region, scaled down from the real 5,700+ sites and 3,900 stores: places, operators, people and changes are dummies. The layers, change rules, plan method and scan checks are the project’s real ones.

Under the hood

Claude
Python
NumPy
SciPy
Shapely
scikit-learn
Playwright
Leaflet
Excel

01Collect (OSINT)
Only the operators’ own published lists: eight public sources, some read in a browser, two by searching map areas twice on shifted tiles. Every raw file is fingerprinted with SHA-256.
02Source register
Where each list comes from, what each field means and the evidence for it, the exceptions, and the rule that a partner’s site counts once, for its own operator.
03The gate
19 conditions stop a scan: a sudden change of more than 10% in a count, missing groups, vanished IDs, swapped coordinates, duplicates, lists cut at a page limit, partner labels that fall away. A failed scan is set aside and never becomes the baseline.
04Change report
New, removed, moved and renamed sites, matched by operator and ID, with changes within 300 m of the client’s own sites flagged. Every removal is checked by hand on the operator’s own map.
05Planning map
Ten views over the sites, an official 1 km population grid, area boundaries and 3,900 chain stores: coverage, client gaps, area scores, overlaps, chains, changes and an expansion planner, all exporting to Excel.
06Verification
A self-test with a deliberately broken source, an identity run that must rebuild last week’s map exactly, and every published figure recalculated three independent ways.

In numbers