Adding a "far from X" factor to a site-selection score can push top zones out of listing coverage; search deeper before falling back
Context: an H3-grid site-selection model gives each zone a weighted score, then searches commercial listings near the best zones (best first, up to a limit).
What happened: I added a distance-from-own-offices factor (straight-line km to the nearest existing office, capped at 15 km) on top of an existing distance-from-competitors factor. Both reward remote cells, so the top 400 zones moved entirely outside the listings database coverage. The search found nothing and the run silently fell back to zone centroids.
Fixes: (1) Search a much larger slice of zones in best-first order. I went from the top 400 to the top 3,000 of about 12k cells. Final ranking is still by score, so going deeper costs little. (2) Run any constraint-relaxation pass before the empty-result fallback, not after it. Otherwise the fallback output counts as a full list and the relaxation never runs. (3) Cap distance features so being far beyond any real effect does not keep adding score.
General lesson: after adding any feature that rewards emptiness, check where the top zones land relative to your candidate inventory.
One gotcha: if your own offices go through the same point-layer ingest as competitors, drop the per-cell point counts after validation. Otherwise your own offices can turn into Huff competitor supply. Also validate the layer on rows read rather than points inside the grid, since a group may have only one office.