Geo Targeting That Survives an Audit of the Bid Stream

In a typical Indian display bid stream, well under a tenth of impressions carry a GPS-derived latitude and longitude. The rest is IP-derived or user-provided, and IP here is weaker than the vendor documentation implies, because Jio and Airtel route very large user populations through small blocks of carrier-grade NAT addresses. PIN-code targeting sold on IP data is, across most of this country, a guess with a confidence interval printed next to it. The audit that comes before any fence is drawn: I read your delivered impressions rather than the plan. The distribution of device.geo.type across your supply. The precision of the coordinates you are paying for, since two decimal places describe a square roughly a kilometre across and fifteen decimal places describe a number somebody generated. The split between app and web inventory. And the share of impressions carrying no location at all that were nonetheless bought as local. That last figure is usually the reason a geo campaign underperformed, and it is knowable in an afternoon. Fences drawn, not typed: A one-kilometre radius around a campus centroid buys the highway beside it and a residential block that will never walk in. Polygons are cut to the building footprint, the entrances and the approach roads, uploaded as GeoJSON or KML to DV360 or The Trade Desk, and paired with hour-of-day rules that reflect when the place is occupied. In dense corridors the fence is drawn against the competitor's frontage too, and I will say plainly when a location is too dense for the signal to separate one building from its neighbour, which in a market street usually means it cannot. Measurement, which is geographic: Visit or enrolment lift is read across geographies: matched control areas chosen on pre-period behaviour, a defined pre-window, and a difference-in-differences estimate with the assumptions stated. Panel-based visitation vendors are a directional cross-check and never the headline number, because their Indian panels are thin enough that a campaign can move the panel without moving the market. The fences I will not draw: Hospital campuses, clinic clusters, diagnostic chains and pharmacy chains. A fence around an oncology centre is an inference about a person's health and I will not build it for a medtech client whatever the local rule currently permits. School gates and any fence designed to reach minors, which rules out a common edtech tactic. Places of worship. Fences around addiction, fertility or mental health services in any form. Where it does work: Travel: airport terminals and their arrival halls, rail termini, hotel clusters for conquesting, and destination fences read at return rather than at arrival. Education: coaching corridors, examination centres on results week, university districts against admission windows. Medtech: conference venues and trade halls where the audience is professional and self-identified rather than patients inferred from a building. Not included: Buying location datasets. Advertising ID sourcing. SDK deployment in your app. Out-of-home planning, which is a different discipline with different measurement and I will refer you rather than pretend otherwise. Store or centre selection, which is your operations team's decision. Who this is not for: Advertisers with a handful of locations and no in-store or enrolment signal to measure against, where the honest answer is that a fence cannot be evaluated and you should spend the money elsewhere. Teams who need a campaign live this week, since the audit precedes the build and the build is where the result comes from.