Metrics & Method

Dwell Time, Capture Rate, Conversion: What Each Retail Metric Really Counts

October 19, 2026 · 12 min read · Teampl Consulting

Three numbers show up on every mall and retail analytics dashboard. They measure three different moments in a shopper's visit, and mixing them up produces bad decisions.

Dwell time, capture rate, and conversion rate are three distinct measurements of retail foot traffic, each built from a different slice of the data a WiFi or BLE sensor can see. Dwell time counts how long a detected device sits inside a defined zone. Capture rate counts what fraction of a larger population — passersby, or connected devices — becomes a smaller, more engaged subset. Conversion rate counts what fraction of visitors complete a transaction. They sound like synonyms in a sales deck. They are not.

The confusion is not accidental. Vendor documentation routinely uses "capture rate" to mean at least three unrelated things, and dwell time gets reported at different spatial resolutions depending on whether the underlying signal is WiFi probe requests or BLE beacon broadcasts. Add Europe's data protection rules — which treat the MAC address at the center of this entire measurement chain as personal data — and the technical and legal pictures become inseparable. This piece works through what each metric actually counts, how the underlying WiFi/BLE capture pipeline produces it, and where GDPR constrains or reshapes the whole exercise, with France, Germany, Switzerland, Belgium and the Netherlands leading the comparison.

None of this is academic. A mall operator comparing tenant performance, or a retail chain deciding where to put a greeter, needs to know whether a reported "capture rate" of 70% describes shoppers entering a storefront or visitors completing a WiFi login form. Those are different problems with different fixes.

Key Takeaways

  • Dwell time, capture rate and conversion rate measure three separate stages of a visit — passing by, entering, and buying — and none substitutes for the others.
  • "Capture rate" is used in the industry for at least three distinct calculations, from passerby-to-visitor ratios (20–35%) to portal-login completion (25–80% depending on incentive) to purchases-per-entry.
  • WiFi probe requests and BLE beacons detect devices, not people, and MAC randomization (standard since iOS 8 and Android 6) means raw device counts overstate unique visitors unless corrected.
  • Under GDPR, a MAC address is personal data per the CJEU's Breyer ruling logic, which means legal basis, anonymization timing and consent design are not optional extras — they are what makes the measurement lawful at all.
  • France's CNIL has published the most detailed operational rulebook in the EU; the Netherlands and other member states apply the same GDPR principles but weigh legitimate interest differently in practice.

What Do WiFi and BLE Sensors Actually Detect?

Every WiFi-enabled phone periodically broadcasts a probe request, a short management frame asking nearby access points whether any known network is in range. WiFi-tracking technology relies on devices such as smart phones sending so called probe requests, broadcast at regular intervals to see which known or unknown wireless networks are available to connect to. A sensor does not need the phone to join a network to see this signal; passive listening is enough. Active and passive scanning have been standardized in IEEE 802.11 to detect access points , which is why almost any commercial access point, not just purpose-built sensors, can be repurposed into a counting device.

BLE beacons work differently. A beacon is a small, battery-powered transmitter that continuously broadcasts an identifier; a nearby phone's Bluetooth radio, or more precisely an app running on it, detects the signal and estimates proximity from received signal strength. Bluetooth Low Energy refers to a low-power wireless protocol designed for intermittent, short-range broadcasts rather than sustained data transfer, which is why it suits beacon-style proximity detection better than general networking. The practical difference matters for compliance: WiFi probe sensing is passive and ambient, picking up any device in range whether the owner has opened an app or not, while most BLE beacon deployments require an app with Bluetooth permissions already granted, folding consent into the app's own permission flow. Our companion piece on probe requests and MAC randomisation goes deeper into the packet-level mechanics for readers who want the full technical trail.

Both methods identify a device by its MAC address, a 48-bit hardware identifier. An example of a MAC address is 00-1F-3F-D7-3C-58, and the MAC address of a WiFi access point is called a BSSID, or Basic Service Set Identifier. Since 2014, phone manufacturers have undercut this identifier's reliability on purpose. Apple added MAC randomization to iOS starting with version 8, Android followed from version 6, and a phone can rotate its broadcast address as many as 20 times during a 40-minute shopping session. A sensor counting raw MAC addresses will therefore overcount unique visitors unless its software applies statistical deduplication. MAC address randomization has made individual tracking unreliable, but statistical methods still produce accurate aggregate footfall data — the trade-off is that individual re-identification across days gets harder, while population-level counts stay usable. Detection without statistical correction for randomization is not a minor rounding error; it is a structural flaw in the resulting dataset.

Dwell Time: What It Counts and What It Misses

Dwell time measures the duration a detected device remains inside a defined geographic zone, usually bounded by one or more access points' detection radius. Vendors typically set a threshold to separate a real visit from a drive-by signal: one common convention counts devices detected for more than two minutes as visitors and those detected briefly, under two minutes, as passersby who likely just walked past . That threshold is a modeling choice, not a law of physics, and two platforms using different cutoffs will report different dwell averages for the same physical store.

Spatial resolution is the bigger limitation. WiFi-based heatmaps are zone-level accurate rather than meter-level accurate — you can tell that one wing of a mall gets three times more traffic than another, but you cannot tell that aisle seven outperforms aisle eight within a single store, and meter-level accuracy needs BLE beacons or camera-based systems instead. Consider a mall operator comparing an anchor-tenant wing against a specialty-retail wing: one wing saw 34% more total foot traffic, but the other wing's average dwell time ran 40% longer, suggesting higher per-visitor engagement despite the lower count . That contrast is only visible because dwell time and raw footfall were reported separately rather than collapsed into one composite score.

Dwell time is not a proxy for intent to purchase on its own. It correlates with engagement, but a long dwell time can also mean a confused shopper standing still at a directory, or a device left stationary in a stroller near the entrance. Reporting dwell time without a zone breakdown — entrance versus fitting room versus checkout queue — is not a stylistic choice but a structural requirement for the number to mean anything operationally.

Why Does Capture Rate Mean Three Different Things?

This is the term that causes the most cross-vendor confusion, because at least three unrelated calculations share the same label. The first usage treats capture rate as a traffic metric: the share of people passing a storefront who actually walk in. Capture rate measures how many people passing a location enter the store , and a typical retail store sees a passerby-to-visitor conversion rate of 20–35% . The second usage treats capture rate as a data-quality metric for a captive WiFi portal: the share of already-connected devices that complete a login or authentication step. A portal capture rate with an incentive attached benchmarks at 65–80% on average across one platform's deployments , while without an incentive the range drops sharply and varies by authentication method. The third usage, seen in some consumer-facing explainers, redefines capture rate as something closer to conversion: capture rate as the percentage of customers entering a store and making a purchase, computed by dividing purchases by the number of customers who entered .

"Capture rate" variantWhat it actually measuresTypical reported range
Traffic/footfall capture ratePassersby who enter the store or venue 20–35%
Portal capture rate (incentivized)Connected devices completing WiFi authentication 65–80%
Portal capture rate (no incentive, by method)Connected devices completing authentication 25–35% email, 18–28% SMS OTP, 38–48% WhatsApp OTP, 15–25% social login
Capture rate as purchase conversionPurchases divided by store entries Entries who make a purchase, not reported as a standard range
Nobody is wrong, exactly — each definition is internally consistent within its source. But reading a vendor report's "capture rate of 68%" without checking which of these four calculations produced it is close to meaningless. Capture rate is not a single metric but a label three or four different measurements share, and the number is only interpretable once its denominator is named.

Conversion Rate: Where WiFi Data Has to Hand Off to POS

Conversion rate in retail refers to the proportion of store visitors who complete a purchase, gauging how well a retailer turns foot traffic into sales. The formula is simple arithmetic — conversion rate measures how many of those visitors buy — but the data pipeline behind it is not. WiFi and BLE sensors see devices entering and dwelling in a zone. They do not see a till receipt. Calculating foot traffic against transaction correlation requires POS integration or manual cross-referencing , because the visitor count and the sales count live in two completely different systems with two different definitions of "the same period" and "the same store."

This is why conversion rate breaks the pattern of the other two metrics: it cannot be computed from WiFi/BLE data alone, no matter how good the sensor placement is. A retailer with excellent dwell-time and capture-rate instrumentation can still have no reliable conversion number if the POS feed is not timestamped and matched to the same visitor-counting window. Decomposing the three metrics together is more diagnostic than any one alone: pairing conversion with capture rate — the share of passing foot traffic a storefront pulls inside — shows that low capture with strong conversion means the business sells well to the few shoppers it attracts, so the real opportunity sits at the window, not on the floor .

Capture rateConversion rateLikely diagnosis
LowHigh Storefront pulls in few shoppers, but converts them well — fix the window, signage, or passing-trade relevance
HighLow Good volume enters, but something inside the store — staffing, layout, stock — is losing them
HighHighStore and storefront are both performing; look at basket size next
LowLowStructural problem — location, category relevance, or pricing, not a quick fix

A team running mystery-visit programmes alongside sensor data gets the ground truth that WiFi/BLE alone cannot supply — what actually happened on the floor between entry and exit, not just that a device was present for eleven minutes.

How Accurate Is WiFi/BLE Detection, Really?

Accuracy depends heavily on sensor placement and the detection method chosen, and the published ranges are wide enough to matter for procurement decisions. A properly configured retail deployment achieves 85–92% device detection accuracy, with access points positioned at entry points and department boundaries providing the most accurate zone-level data. After statistical correction for MAC randomization, aggregate counts typically land within a tighter but still real error band: WiFi presence analytics runs roughly ±10–20% after deduplication, camera-based systems run closer to ±3–5%, and hardware-native positioning platforms can hit sub-3-meter location accuracy while still carrying about ±15% error on raw footfall counts .

Detection methodRelative costReported accuracySpatial resolution
WiFi presence (probe requests)Low — uses existing APs ±10–20% after dedup Zone-level
Camera-based countingHigher — dedicated hardware ±3–5% Sub-meter
Hardware-native positioning (e.g. enterprise WiFi platforms)High — vendor-specific infrastructure ±15% footfall, sub-3m location Meter-level
BLE beaconsLow-to-moderate — beacon hardware plus appVaries by app adoptionMeter-level where deployed
In mall settings, detection accuracy also depends on corridor geometry: with access points positioned at corridor junctions, food court boundaries, and entrance/exit points, detection rates reach 80–90% of all visitors carrying WiFi-enabled devices . One of Teampl's earliest field data programs was large-scale WiFi signal-scanning, run across Germany, France and the Benelux region at a throughput of thousands of scans daily — the kind of ground-truth survey that checks whether a planned access-point layout actually matches real-world signal propagation before a mall-wide analytics rollout goes live. Published accuracy benchmarks describe what a well-placed sensor network can achieve; verifying that a specific mall's AP layout meets that standard is a separate, physical exercise.

Does GDPR Allow This at All?

Yes, but only within a fairly narrow set of conditions, and the EU — led operationally by France — has the most developed rulebook anywhere. The starting point is settled: a MAC address counts as personal data. As the GDPR considers MAC addresses pseudonymised personal data capable of singling out a person, their use requires a valid legal basis and adherence to the regulation's other articles , a conclusion that follows the CJEU's reasoning in the 2016 Breyer case on dynamic IP addresses. The foundational EU guidance on this predates GDPR itself: the Article 29 Working Party's Opinion 13/2011 on geolocation services concluded that prior informed consent is the main applicable ground for processing location data derived from base station, WiFi and GPS signals by information society service providers , and that logic carried straight through into GDPR-era enforcement.

France's CNIL has published the most granular operational guidance in the bloc. Its October 2018 rules set out three compliant scenarios for audience-measurement sensors in shopping malls and similar venues. Where data is anonymized at short notice, the CNIL recommends completing that anonymization within five minutes of collection, after which no identifying data should be retained. Where pseudonymized data is retained longer, the operator must implement a reliable pseudonymization process using a salt or key, delete the raw MAC address, and anonymize or delete the pseudonymized data by end of day. Crucially, pseudonymization alone does not equal anonymization in the CNIL's view: anonymization must satisfy the former Article 29 Working Party's Opinion 05/2014 criteria, including ensuring a high collision rate so that multiple MAC addresses map to the same processed identifier. That distinction killed a real pilot. In the JCDecaux/Fidzup pedestrian-tracking case on the La Défense esplanade, the CNIL found that hashing MAC addresses without sufficient collision was a pseudonymisation technique, not an anonymisation one, since a given identifier typically still matched only one person. Where operators want to let visitors opt out directly, the CNIL recommended a dedicated WiFi network with a clear, explicit name such as "wifi_tracking_optout" that automatically collects the MAC address of anyone exercising their right to object. Immediate, high-collision anonymization is not a stylistic choice in France; it is the structural requirement that separates a lawful audience-measurement sensor from an unlawful tracking device.

The Netherlands applies the same regulation but with a somewhat different weighing exercise. Dutch analysis has accepted legitimate interest as a workable legal basis for aggregate counting in some circumstances: a lessor of retail space may have a legitimate interest in collecting business-economic information about visitor numbers per store over time, though that interest must be weighed against the interests of shop visitors and passers-by. Dutch commentary has also pushed back on the idea that public versus private space changes the analysis much: the distinction between privately and publicly owned space looks rather arbitrary, since train stations and enclosed shopping malls are usually privately owned too, which makes it hard to argue those parties automatically carry a legitimate interest that others don't. A real example from Dutch Railways shows the pattern in production: the operator relies on legitimate interest — "to improve our services and to increase your safety in and around the station" — and immediately hashes the MAC address, using a different hash per day so data cannot be correlated across multiple days.

Germany, Belgium and Switzerland apply materially similar principles without France's level of published sensor-specific guidance. Germany's state-level data protection authorities assess WiFi and BLE tracking under the same GDPR legitimate-interest balancing test, generally expecting the same anonymization-at-source approach the CNIL spelled out. Belgium's Data Protection Authority follows EDPB-aligned guidance rather than issuing a dedicated sensor framework. Switzerland sits outside GDPR's direct territorial scope but applies its own revised Federal Act on Data Protection, which imposes comparable consent and minimization obligations and is the subject of an EU adequacy arrangement, so a multinational mall operator running sensors across all four countries in practice converges on one compliance design rather than four.

MarketTypical legal basis usedAnonymization expectationKey reference point
FranceLegitimate interest or consent, by scenario Within 5 minutes, high collision rate CNIL 2018 guidance; Art. 29 WP Opinion 05/2014
NetherlandsLegitimate interest (weighed case by case) Salt rotated daily Legitimate-interest analysis for lessors
Germany / BelgiumGDPR legitimate interest, EDPB-alignedAnonymization at source expectedState/national DPA guidance, no dedicated sensor framework
SwitzerlandRevised FADP (non-GDPR, materially similar)Comparable minimization obligationFADP 2023 revision
United StatesSelf-regulatory code or state opt-out regimesNo standardized timing requirementFPF Mobile Location Analytics Code of Conduct
Outside the EU, the regulatory posture is noticeably looser. The US has historically relied on industry self-regulation rather than a statutory opt-in requirement: the Mobile Location Analytics Code of Conduct, a self-regulatory framework published by the Future of Privacy Forum, has governed much of this space rather than a GDPR-style consent mandate, and state privacy laws generally give consumers an opt-out right rather than requiring opt-in consent before collection starts. That gap is why a global retailer cannot run one compliance playbook across all markets — the EU design has to travel upward in strictness, not the other way around.

What Does a Compliant Deployment Actually Look Like Day to Day?

Beyond the legal basis question, a working compliance layer has concrete engineering requirements. If the WiFi analytics involve large-scale profiling or location tracking, a data protection impact assessment documenting the privacy risks and mitigations is typically required. The compliance layer sits alongside, not instead of, the analytics layer: consent management, data retention policies, anonymization and privacy controls are what keep the data collection within GDPR and CCPA boundaries. And some data that a sensor is technically capable of capturing should simply never be collected: WiFi analytics can technically capture browsing history, DNS queries and application usage from connected devices, and that data should not be collected. The gap between what a sensor can see and what a lawful program should record is itself a design decision, documented in the DPIA rather than left implicit in a vendor's default configuration.

What's a Realistic Worked Example?

Consider a mid-size retail store instrumenting a captive WiFi portal with an incentive-based opt-in. A store with 400 daily visitors and a WiFi captive portal capturing a 20% opt-in rate builds a database of 80 identified shoppers per day — 2,400 per month. Layer in passive presence analytics alongside the portal, and the same store can report total footfall (everyone detected, portal login or not), capture rate (what fraction of footfall becomes an identified profile), dwell time by zone, and — only once POS data is matched to the same time window — conversion rate. Four numbers, four different data sources feeding them, and four different failure modes if any one pipeline breaks silently. A mall operator running this across dozens of tenants faces the same reconciliation problem at a larger scale, which is part of why structured, field-verified ground truth — the kind documented in our piece on what mystery shopping measures and cannot — remains a necessary complement to sensor data rather than a redundant one.

Where Public Benchmarks Fall Short

Published benchmark ranges are genuinely useful for sanity-checking a vendor's numbers, but they describe averages across unnamed populations of stores, not a specific mall's specific tenant mix. A reported 4.7% annual decline in shopping-center foot traffic from 2019 through 2024, which stabilized in 2025, with analytics-enabled malls outperforming the average by 12% in visitor retention tells an operator what the industry trend looked like, not whether their own property is above or below it this quarter. Benchmark figures are a starting hypothesis to test against a property's own instrumented data, not a substitute for it.

Want the operational detail behind how programs like this actually get run — recruiting, quality control, chain of custody? That's what we do day to day.

Frequently Asked Questions

Is a hashed MAC address still personal data under GDPR?

Usually yes, unless the hashing meets strict anonymization criteria. Anonymization must satisfy the Article 29 Working Party's Opinion 05/2014 criteria, including a high collision rate between individuals, and France's CNIL recommends completing this within five minutes of collection. A salted hash that still maps one MAC address to one person is pseudonymisation, which keeps the data inside GDPR's scope, not outside it.

Why do two vendors report different dwell times for the same store?

Because dwell time depends on a threshold choice — how many minutes of detection count as a "visit" versus a pass-by — and on access point placement, which determines the zone boundary being measured. One common convention treats devices detected for more than two minutes as visitors and shorter detections as passersby , but that cutoff is a vendor configuration choice, not a regulated standard.

Can WiFi/BLE data alone calculate conversion rate?

No. Foot-traffic-to-transaction correlation requires POS integration or manual cross-referencing , because sensors see device presence, not completed purchases. Conversion rate is the one metric of the three that structurally depends on a second data system beyond the sensor network.

Next step

Want the operational detail behind how programs like this actually get run — recruiting, quality control, chain of custody? That's what we do day to day.

Start a project