How Accurate Is Foot Traffic Data? A Buyer's Methods Guide
Foot traffic data is an estimate, not a count. Vendors watch a sample of phones, assign their location pings to places, and project the result to the population. It can track chain-level trends well. Single-store, single-week and multi-tenant reads are much weaker. Accuracy depends on panel bias, GPS error, polygon quality and how the vendor validates.
How to use this article
Read the checklist here, then use the linked hub and product pages for procurement citations.
No vendor counts every person who walks into a store. What they have is a panel: phones running apps that share location, usually through an SDK, with the user's permission. The vendor cleans those location pings, decides which ones add up to a visit to a particular place, and counts visits among panel devices. That count is then scaled up. A simple version weights each device by how many people it stands for in its home area, based on census population and the panel's share of it. A more careful version also corrects for the panel growing or shrinking over time, so a new app partner does not look like a surge of shoppers.
This means two separate things can go wrong. The observed visits can be wrong, because a ping was assigned to the wrong place or a drive-by was counted as a visit. And the projection can be wrong, because the panel does not look like the people who actually shop there. A vendor can be good at one and weak at the other, so ask about both. For how many devices a read needs before it means anything, see foot traffic panel sizing.
Device Sampling Bias: Who the Panel Leaves Out
A panel can only see people who carry a smartphone, run an app that shares location, and allow it. Each step filters the population. Pew Research Center's mobile fact sheet shows that smartphone ownership is lower among adults 65 and older than among younger adults, so older shoppers start underrepresented before any app is involved. Platform matters too. Apple's user privacy guidance for developers requires apps to get the user's permission through the App Tracking Transparency framework before tracking them or accessing the device's advertising identifier, which changes how many iPhone users a panel keyed on advertising IDs can see. And every panel inherits the audience of the apps behind it, whether those are weather, navigation, games or coupons.
The best public measurement of this is an academic audit. In "Leveraging Administrative Data for Bias Audits", presented at the ACM FAccT conference in 2021, Amanda Coston and co-authors compared SafeGraph's mobility data with official records of in-person turnout at polling places in North Carolina's 2018 general election. They estimated that traffic to a location with a younger, largely white visitor population was more than four times as likely to be represented in the data as traffic to a location with older, non-white visitors. That result is about one dataset at one point in time, and it is not a verdict on any vendor today. What it shows is that the skew is large enough to matter and that it can be measured when a real ground truth exists. Ask your vendor what it has measured and how it reweights.
GPS Horizontal Accuracy and Why Dwell Matters
A location ping is not a point. It is a best guess with an error radius. According to the archived GPS.gov accuracy page, GPS-enabled smartphones are typically accurate to within a 4.9 meter (16 foot) radius under open sky, and accuracy gets worse near buildings, bridges and trees. Most shopping happens next to or inside buildings, which is exactly where that figure stops holding.
Phones report that error with each fix, and the definition is worth reading closely. Android's Location.getAccuracy) returns the radius of 68% confidence: there is a 68% chance the true position is inside that circle. So roughly one fix in three sits outside its own stated radius. A vendor that keeps a ping reporting 50 meters of accuracy next to a storefront 20 meters wide is guessing which business the phone was in.
This is where dwell time comes in. One ping inside a shape could be someone walking past, driving through the parking lot or waiting at a light. Several pings over a few minutes, with good accuracy, inside the same shape, look much more like a visit. Dwell thresholds trade recall for precision: set them short and you count passers-by, set them long and you drop the quick coffee run. Neither setting is correct in general, which is why a vendor should state the threshold it uses per category. The MRC Location-Based Advertising Measurement Guidelines, written with the IAB and the MMA, expect measurement vendors to disclose their methodology and to disclose validated location precision and accuracy information in their reports. That is a reasonable standard to hold any vendor to.
The shape matters as much as the ping. A circle around a pin at the center of a lot counts everyone parked nearby. A polygon traced around the building footprint does not. Our guide to polygons vs radius geofencing covers that tradeoff, and our post on POI data quality and foot traffic accuracy covers the place side in depth: centroid errors, brand hierarchy, category tags and closures.
Visit Attribution in Dense and Multi-Tenant Buildings
Attribution is hardest where businesses are packed together. In a mall, a strip center or a downtown block, the error radius of a single ping can cover several tenants at once. In a building with a pharmacy on the ground floor and offices above, most location data has no reliable floor, so a visit to the dentist upstairs and a visit to the pharmacy downstairs can look the same. Shared entrances and shared parking make it worse.
Vendors handle this in different ways. Some assign the visit to the nearest polygon. Some split it by probability across the candidates. Some drop ambiguous visits altogether. Each choice biases results differently. Nearest-polygon inflates the anchor store. Probability splits smear visits across neighbors. Dropping visits undercounts dense locations compared with standalone ones. None of these is wrong, but they are not interchangeable, and they matter most when you compare a mall store to a freestanding one. Ask which method the vendor uses and whether it flags visits it considered ambiguous. For the place-data side of the same problem, see the same address, different business problem in our POI quality guide.
How Vendors Validate Against Ground Truth
Validation means comparing the estimate to something counted independently. The best ground truth is a retailer's own data, such as door counters, transactions or loyalty records, shared under agreement. Vendors rarely publish those comparisons because the retailer owns the numbers. What is public is mostly comparison with reported company results.
The academic version is a working paper published by the National Bureau of Economic Research. In "What Do Measures of Real-Time Corporate Sales Tell Us About Earnings Surprises and Post-Announcement Returns?", Kenneth Froot, Namho Kang, Gideon Ozik and Ronnie Sadka built real-time proxies of retail sales from sources including roughly 50 million mobile devices, and found that the measure helped explain quarterly sales growth and revenue and earnings surprises. The best-known public example came in April 2016, when Fortune reported that Foursquare had used location data from Chipotle's more than 1,900 US restaurants to forecast a first-quarter same-store sales drop of about 29%. Chipotle then reported that comparable restaurant sales fell 29.7%.
Read those results for what they are. They are chain-level and quarterly, and they cover public companies with large store counts, which is where a panel is strongest. A forecast that matched once may be the one that got published. They show that foot traffic can track a real outcome when enough stores and weeks are averaged. They do not show that a single store's weekly count is right. If you mainly need chain trends, see foot traffic vs credit card panels for when a spend panel is the better check.
Questions to Ask a Foot Traffic Vendor
How many devices were observed at my locations in the period, not in the national panel?
How do you project from panel visits to total visits, and how do you correct for panel size changing over time?
What demographic or platform skew have you measured in the panel, and against what ground truth?
What horizontal accuracy cutoff do you apply to pings, and what share of pings does it remove?
What dwell threshold defines a visit, and does it change by category?
How do you attribute visits in malls and multi-tenant buildings, and can you flag ambiguous visits?
Which validation studies can you share, at what level (chain, store, week), and with what error measure?
Can you run a test on locations where I hold my own counts before I sign?
Accuracy Claims to Distrust
A single accuracy percentage with no definition. Accuracy of what: device location, visit detection, projected totals or trend direction?
Correlation with revenue quoted without the level. A high correlation across a year of chain totals says nothing about one store in one week.
Validation on a hand-picked case. Ask how many chains or quarters were tested and how many missed.
Panel size quoted nationally. What matters is devices seen at your locations.
Precision to the meter. A reported coordinate with six decimal places does not make the underlying fix any better.
Claims that the data counts everyone. It is a sample, and a vendor that says otherwise is not describing its method.
It depends on the level. Foot traffic data is a projection from a sample of phones, so chain-level trends over weeks or quarters are usually its strongest use, and public studies have found it tracks reported sales at that level. Single-store, single-week and multi-tenant reads carry much more error from GPS drift, attribution rules and small samples.
How accurate is smartphone GPS for store visits?
The archived GPS.gov accuracy page says GPS-enabled smartphones are typically accurate to within 4.9 meters under open sky and worse near buildings, bridges and trees. Android reports accuracy as a 68% confidence radius, so about one fix in three is outside its own stated circle. That is why vendors filter pings by accuracy and require dwell time before counting a visit.
Is foot traffic data biased?
Every panel is. It only sees people who carry a smartphone, use an app that shares location and allow it. A 2021 academic audit estimated that one mobility dataset was more than four times as likely to capture traffic at places with younger, largely white visitors than at places with older, non-white visitors. Ask vendors what skew they have measured and how they reweight.
How do vendors validate foot traffic data?
The strongest check is comparison with a retailer's own counts, such as door counters or transactions, which is rarely published. Public checks mostly compare estimates with reported company results, such as the 2016 case where a location-based forecast of Chipotle's sales drop closely matched the 29.7% decline Chipotle reported. Those checks are chain-level and quarterly.
Why is foot traffic less accurate in malls and office towers?
Businesses sit close together or stacked on different floors, and a single location ping's error radius can cover several of them. Most location data has no reliable floor. Vendors assign these visits to the nearest place, split them by probability, or drop them, and each choice biases dense locations differently from standalone stores.
✓ Opt-Out Request Honored via Global Privacy Control
We use cookies and similar technologies to improve your experience. No marketing cookies are pre-selected. Learn more in our Privacy Policy. Your Privacy Choices.
