Why POI Data Quality Makes or Breaks Foot Traffic Analytics

Most foot-traffic failures are POI failures, not panel failures. When the underlying database of places has outdated addresses, missing locations, collapsed multi-tenant addresses, or sloppy boundaries, even a strong device-level mobility panel produces misleading visit counts: wrong geofence, wrong visit count, wrong lift, share, and site score. GSDSI's POI data and POI & Geofencing exist because the polygon-and-taxonomy layer is the silent determinant of every downstream read, from CPG competitive benchmarks and CRE site selection to CTV-to-store attribution. This guide explains what POI data quality means and then gives the accuracy tests a buyer should run before licensing. For shortlisting vendors, pair it with the POI data provider comparison, foot-traffic panel sizing, and geofencing best practices.

How accurate is POI data for ad targeting?

Polygons vs. Radius: The 30-40% Problem

A 50-meter radius around a Starbucks in a strip mall captures devices visiting the dry cleaner next door, the parking lot behind it, and apartments across the street. A polygon tracing the building footprint (GeoJSON Polygon or MultiPolygon, plus drive-through lanes where relevant) eliminates most of that noise. Real-world migrations from radius to polygon attribution typically cut false-positive visit rates by 30-40% overnight, not because devices changed, but because the join key became accurate. Centroid plus a 50-150m radius is cheap and still acceptable for isolated big-box sites with large lots; it breaks in multi-tenant towers, strip malls, and dense blocks. POI & Geofencing ships polygon-primary records with a centroid fallback per use case.

Procurement should require polygon provenance: parcel data, imagery validation, and manual QA samples in dense corridors. Centroid-only records are a warning sign for any use case requiring venue-level attribution. Document the geometry type in every benchmark, because a year-over-year read cannot be audited if a Q3 radius quietly became a Q4 polygon.

Where Centroid Placement Errors Hide

Where a centroid is still used, its placement matters. The geometric center of a building can land on the rooftop while the entrance faces the parking lot, and a 30-50m placement error is enough to misattribute a neighbor's traffic. Ask each vendor whether the point is a footprint center, an entrance, or a parcel center. Entrance-aware centroids with a polygon override for the visit check is the right architecture.

The "Same Address, Different Business" Problem

Consider 123 Main Street with a coffee shop on the ground floor, a dentist on the second, and offices above. An address-level record collapses all three and mixes customer types. The fix is tenant polygon cut-outs with a NAICS primary code per tenant, which requires survey-grade sourcing or a parcel and tax crosswalk. It is expensive and it is mandatory for dense urban foot-traffic. Without unit IDs inside the parent polygon, models attribute every visit to the anchor tenant or to the parent property record, hiding true tenant performance.

• Coffee (NAICS 722515) is not an office (NAICS 523930); filter downstream by code.
• Malls and mixed-use sites without unit boundaries are unusable for tenant-level reads.
• Verify the vendor's method in the pilot on about 50 urban sites.

Brand Hierarchy: Why Rollups Break Silently

If your database lists Taco Bell and KFC separately without linking them to Yum! Brands, roll-up analysis for investment research or competitive benchmarking will be incomplete. The same applies to franchise versus corporate locations, co-branded stores, and ghost kitchens under virtual brand names. Analysts using alternative data in equity research notice immediately: the same brand can look present or absent depending on corporate parentage graphs.

• Location, brand, parent brand, and corporate entity chain.
• Franchise flags and co-tenancy markers for shared addresses.
• Virtual brand aliases mapped to operating kitchen POIs.

NAICS Tagging and Category Comparisons

Every POI should carry standardized industry classification mapping to NAICS at a consistent grain. Comparing QSR versus fast-casual foot-traffic trends in a DMA only works with accurate category assignments. Enterprise POI files should conform to or crosswalk against NAICS at the 4- or 6-digit level, aligned with BLS QCEW reference data where stakeholders expect public-sector reproducibility. Category errors are worse than missing locations because they poison benchmarks while looking precise, so request a sample manually reviewed against NAICS for your target vertical before production.

Refresh Cadence and the Closure Problem

The U.S. Census Business Dynamics Statistics program documents establishment churn at national scale: roughly 7-9% of retail establishments close each year, with a similar share opening. On an annual refresh, that share of records is stale within 12 months and 13-17% within 24, which means false visits to closed stores and missed new openings. Monthly refresh with open and close dates on every record is the working minimum, and sub-72-hour retirement of closed locations once a detection signal is available is the defensible target for active measurement programs. Require a change-delta covering a shutdown or rebrand week you already verified; a vendor promise is not evidence.

How to Evaluate POI Accuracy: Build a Ground-Truth Venue List First

Pick 50 to 100 venues you operate or can verify: airports, malls, downtown flagships, suburban boxes, and medical plazas if relevant. Archive map screenshots. Vendors should match that list, not a marketing NAICS rollup. Weight urban strip-mall and multi-tenant cases higher if your portfolio is dense-market heavy, because suburban-only pilots hide the failures that break attribution, and publish weights before scores so incumbents cannot reframe after the fact.

Run POI QA before any mobility bake-off. Score POI and mobility vendors separately, since a winning mobility panel on bad POIs will lose in production. Agree visit definitions in writing (dwell, employees, multi-visit same day) before comparing vendors: POI fixes visit location, but mobility vendors still differ on visit logic, and disagreement after a polygon migration usually means visit logic differed rather than that polygons failed. Where possible, compare attributed visits to operator-reported traffic. For how panel bias, GPS error and dwell rules shape those visit counts, see how accurate foot traffic data is.

Auditing Polygon Fidelity Without a GIS Team

Vendors describe polygon quality in adjectives. You can measure it in an afternoon with a spreadsheet and public imagery. Draw a stratified sample of about a hundred POIs, weighted toward the venue types that actually matter to your analysis, then compare each delivered footprint against satellite imagery and, where available, the parcel boundary. Score three things per record: whether the footprint covers the building the business occupies, whether it spills into neighboring tenants, and whether it includes outdoor area no customer would ever stand in.

Express the result as an overlap ratio rather than a pass or fail. A footprint covering 70 percent of the true premises and nothing else is far more useful than one covering 100 percent plus the two units either side, because the first under-counts predictably while the second imports a competitor's customers into your visit series. Predictable under-counting can be calibrated. Contamination cannot, because you cannot tell afterwards which visits were yours.

The Two Failures a Ground-Truth Match Cannot Catch

A ground-truth list answers one question well, which is whether the vendor has your venue and drew it correctly. It cannot see the errors that live outside the list. The first is duplication. Query each ground-truth venue by name and by geography rather than by the vendor's ID, then count how many records come back. A venue carried twice splits visits across two IDs, and every count for that location is understated in a way a missing-venue check never reports. Ask how the vendor deduplicates across upstream sources, what breaks a tie when two sources disagree on geometry, and whether a retired duplicate keeps a pointer to the surviving ID so older extracts still join.

The second is opening latency, the mirror of the closure test. New venues are harder to audit because you cannot verify what you did not know existed, so build the test from openings you can date yourself, such as your own new locations or a competitor's announced sites. Compare first appearance in the file against the actual opening date. Ask as well whether a venue arrives complete or lands as a bare point and collects its polygon, hours, and category later. A record that is present but geometrically unfinished passes a match test and behaves like a miss in attribution. Both errors flatter the vendor, because both understate visits at exactly the locations a campaign is trying to measure.

Vertical Venues and the Third Dimension Most POI Files Ignore

Almost every commercial POI file is two-dimensional, and consumer GPS is weakest exactly where that assumption breaks. In an enclosed multi-level mall, a medical tower, or a mixed-use block with retail below offices, several businesses share one ground footprint at different elevations. A flat polygon cannot separate them, and horizontal accuracy indoors is often worse than the distance between the units you are trying to tell apart.

Treat vertical venues as a known blind spot rather than a solved problem. Ask the vendor how stacked tenancies are handled and expect one of three honest answers: the visit is attributed to the dominant tenant, it is split by a modeled share, or the location is excluded. All three are defensible. What is not defensible is a file reporting confident per-tenant counts inside a shopping center without saying which of the three it did. If a material share of your locations are vertical, hold them out of headline metrics and report them separately. Retail media networks should validate mall unit boundaries in their top fifty centers before licensing national POI.

Parking Lots, Shared Entrances, and Dwell Contamination

The area around a store creates more measurement error than the store itself. A polygon drawn generously enough to catch the entrance usually catches part of the parking lot, and a lot shared with four other retailers turns every neighbor's customer into a short dwell on your record. Filtering by dwell time helps, but it is a blunt instrument: it also removes genuine quick-visit traffic, which for convenience and quick-service formats is most of the business.

The better diagnostic is the shape of the dwell distribution rather than its mean. A clean site shows a bimodal pattern, a short pass-through cluster and a longer genuine-visit cluster, and the cut between them can be read straight off the data. A contaminated site shows one smeared distribution with no clean separation, which tells you the polygon is wrong before any threshold is chosen. Run that check per location, not per chain, because contamination is a property of the specific site plan.

Coverage Bias Between Chains and Independents

POI coverage is rarely uniform, and the bias runs the same direction almost everywhere. National chains are well represented because their locations are published, branded consistently, and easy to reconcile. Independents, franchisees trading under local names, and businesses operating inside other businesses are thinner, later, and likelier to carry stale open or closed status. If your analysis compares a chain against its independent competition, that gap is not background nuisance. It is the finding, and it will read as the chain gaining share.

Quantify it before modeling anything. Take a handful of representative trade areas, build a ground-truth list of every business present from a second source, then compute match rate for chains and non-chains separately. A twenty-point gap between those two numbers is common and entirely workable, provided every downstream comparison is restricted to the covered set rather than presented as a picture of the whole market.

How GSDSI Builds and Maintains POI Coverage

GSDSI's POI database uses multi-source verification: commercial registries, satellite imagery, web-scraped operational metadata, and on-the-ground validation signals. Records carry polygon boundaries (not just centroids), full brand hierarchy mapping, NAICS-conformant tags, operational metadata (hours, seasonal closures, co-tenancy), and refresh cadences targeting sub-72-hour retirement of closed locations once detectable. Real Estate Data anchors the property layer where governance requires it.

• Polygon boundaries validated against imagery where parcel data exists.
• Brand hierarchy through corporate parentage graphs.
• NAICS tags at 6-digit grain with documented crosswalks.
• Closure detection from imagery, web status, and visit-pattern anomalies.

POI Procurement Diagnostics

1. Share of polygon-primary versus centroid-radius records: below 70% polygon-primary is legacy for dense retail.
2. Centroid placement methodology: entrance-aware is required.
3. Multi-tenant handling: tenant cut-outs versus single-address collapse.
4. Refresh cadence and open or close provenance: monthly or better.
5. NAICS coverage rate: below 95% primary-code coverage leaves gaps in categorical filters.

"Comprehensive" marketing often means address-level, centroid-only, annual refresh. Fail that file in the pilot, not in production. Score vendors with the POI data provider comparison, start from the location intelligence hub when the program spans POI plus mobility, and scope competitive benchmarking for share-of-visit programs. Side-by-side comparisons help executives, but keep numeric scores in a spreadsheet everyone can replay. For end-to-end context, see cross-channel measurement for privacy-first advertisers and the retail site selection stack.

Re-Check After Signing, Because Evaluation Files Are Not Production Files

Accuracy work concentrates on the evaluation sample, and the file that arrives in production is built by a different process at a different time. Evaluation extracts are often assembled with more care, for markets chosen by the vendor, and closer to a manual review than the automated pipeline that will deliver every month afterwards. Re-run a reduced version of your acceptance test on the first two or three production deliveries and then periodically, using the same scoring so the numbers are comparable. Agree in advance what happens if production falls materially short of the evaluated sample, since a remedy negotiated before signing is a contract term and one raised afterwards is a complaint.

Name an owner for the acceptance test who is not accountable for closing the deal, because thresholds tend to soften as a deadline approaches. Put the duplicate query and the opening-latency check into the annual retest alongside the venue list, and store the vendor's POI version ID with every foot-traffic extract and every retest result. When marketing asks why March diverged from February, you need to know whether the panel moved or the polygon changed; without that version stamp, the question becomes an argument neither side can settle. To spot-check per-chain foot-traffic reads before a QA engagement, try the free foot-traffic tools, then request a chain-level sample through contact.