Local Falcon

Local search has no single answer. A business ranks differently a street away than it does outside its own front door, so measuring visibility means sampling results across a grid of real geographic points rather than asking once and recording a number.

That turns a reporting product into a high-volume data problem: many points, many keywords, many locations, repeated over time so trends mean something. The infrastructure behind it holds geographic records in the tens of millions.

Visibility used to mean the map pack. It now also means what a generative AI assistant says when asked for a recommendation, which is a different surface with different behaviour and no ranking positions to read. Tracking it means treating each surface as its own source rather than assuming one model of "rank" covers all of them.

A reporting platform’s value is that its numbers can be trusted. That makes upstream failure a correctness problem, not a availability problem: if a source cannot be reached, the honest answer is to say so.

The lookup and fallback behaviour reports an unavailable source as unavailable. It does not substitute placeholder content, carry forward a previous value as though it were fresh, or quietly produce a report that looks complete and is not.

A grid of positions is evidence, not advice. The analysis service reads completed scans and produces recommendations: what the visibility pattern indicates, which locations and keywords are underperforming, and what could be changed. It runs as its own backend service rather than inside the reporting application.

Beyond reporting, an agent surface lets customers operate the product conversationally: run and manage scanning campaigns, read reports, and manage the Google Business Profile information that visibility depends on. It is built on the same agent platform pattern used elsewhere in the engagement — a runtime with a declared tool surface over real APIs.

At this size the ordinary approaches stop working. Lookups have to be indexed for the access patterns that actually occur rather than for the ones the schema suggests, processing has to be batched, and anything per-record that touches a network becomes the dominant cost of the whole operation.