Recon AnalyticsSurvey data quality is a constant fight - respondents speed through surveys or give nonsensical answers just to collect an incentive, and that noise flows straight into the analysis if nothing catches it. I set out to build an automated way to flag implausible responses before they ever reached our reporting, without relying on personally identifiable location data.
I used the FCC's broadband availability data, which is published using H3, the open-source hierarchical hexagonal spatial index originally built by Uber for location analytics. At the resolution the FCC releases wireless and broadband data, that's over 80 million hexagons across the continental US. I worked out a methodology to rationalize those hexagons down to zip code, the most granular location detail we can use without touching PII, then mapped each zip code to its plausible broadband providers. Every survey response gets checked against that table: if a respondent claims a broadband provider that doesn't actually serve their zip code, it's flagged as low quality. Since there are over 3,000 broadband providers nationally and we can't list them all as fixed survey options, we also let respondents write in their provider as a verbatim, then used AI to match those free-text entries against the plausible provider list for their zip code. That gave us a long-tail view into smaller regional providers beyond the usual top 10-20 names, and another layer to filter on. Once the initial zip-to-hex methodology was built, refreshing it only takes updating with new FCC source data every six months, and the whole check runs automatically as part of our normal ETL process.
The result was a straightforward, automated way to catch up to 20% of raw respondents giving bad answers, all without adding manual review steps. Because of our contractual terms with our survey provider, being able to point to specific, provably erroneous responses meant we could push back on paying for that bad sample, lowering our effective cost per usable response. Higher data quality and lower sample cost, both baked into the standard pipeline instead of requiring extra work.
