Journal · 28 May 2026

Geofence overlap in dense grids

In a city drawn as stacked circles, a single ping can be a mall visit, a store visit, and a lounge visit at once. That is not richness. That is double counting waiting for a slide.

Sukhumvit is the example we keep using because students recognise it immediately. A mall footprint, a wing fence, a brand circle, a competitor circle, and a “VIP” radius drawn by a well-meaning agency can all contain the same coffee table. If your warehouse increments a visit for every matching fence, your weekly report is an argument among polygons.

Overlap is a product decision

Engineers often ask which algorithm should resolve ties. The honest answer is: write the business rule first. Does a lounge visit require an already-confirmed store visit? Does a mall visit count if the user never left the BTS hall that the mall polygon accidentally includes? Those are not clustering hyperparameters.

We teach a simple priority stack: most specific fence wins, but only if the parent fence is also true. Lounge ⊂ store wing ⊂ mall. A ping that matches lounge and not mall is treated as geometry error, not as a VIP appearance.

How inflation shows up

Visit volume that swings when a neighbouring brand redraws a circle. Conversion rates that beat receipt data. “Unique visitors” that exceed gate counts. If any of those appear, print the fence map at A3 and sit with ops before you open Looker.

What not to do

Do not average overlapping fences into a blended score. Do not silently drop the larger fence. Do not let marketing draw circles in a vendor UI without a review that includes the analyst who owns the funnel. The last one is how Rama IX case studies are born; we published one on the reviews page.

Related: dwell is not intent, and the geometry surface on location intelligence.