Linking video analysis to point-of-sale data involves matching two sets of time-stamped events: on one hand, customer visits and dwell times measured by zone; on the other, transactions. The matching is based on time and zone, never on the customer’s identity. It is this rule that makes the process both actionable and compliant.
The expected result is a conversion rate by zone: the percentage of visitors present in front of a shelf, display, or service station who subsequently make a corresponding purchase. Three conditions determine the reliability of the calculation: clocks synchronized between the two systems, a zone division that remains stable over time, and visitor counting that excludes staff.
Taken separately, both systems leave a gap in the data. The checkout system shows what was sold but does not know how many people saw the product without buying it. Video counting shows how many people passed by but does not know what was sold. Cross-referencing fills this gap and allows for answers to specific operational questions: a department attracts few visitors or has low conversion rates; a promotion generated traffic but no sales; a new store location alters foot traffic in neighboring areas.
It also allows for comparing two retail locations on a common basis, provided the data is normalized by total foot traffic and store format.
On the video side, four metrics are sufficient in most cases: the number of visitors who entered the store, the number of visitors who entered each zone, the average time spent in each zone, and the timestamp for each visit, down to the minute.
On the checkout side, you need timestamped receipts, receipt lines with product SKUs, and the correspondence between each SKU and the zone where it is located. This last point is often the most labor-intensive: it requires an up-to-date store layout and a cross-reference table between product families and zones. Without this table, the cross-referencing makes no sense.
The two systems rarely share the same clock. A discrepancy of just a few minutes is enough to assign a transaction to the wrong time slot and make the graphs inconsistent. The first step, therefore, is to synchronize the video recorder and the point-of-sale system to the same time source, then periodically verify that no drift has occurred.
There remains the time lag inherent in the customer’s journey: a customer walks past a shelf and then pays several minutes later. It is therefore necessary to define an attribution window—that is, the duration during which a transaction is linked to a visit to that area. This window is calibrated by observing the actual distribution of travel times between the zone and the checkout in the specific store. It cannot be directly applied from one store layout to another.
The division determines everything else. Three principles emerge from practical experience. A zone must correspond to an actual operational decision—a department that can be reorganized, or a product offering that can be modified. It must be covered by a usable camera view, without any major blind spots. It must be stable enough over time so that comparisons between periods remain meaningful.
A division that is too fine results in low customer counts per zone and per time slot, leading to highly unstable rates. A division that is too broad dilutes the differences and does not lead to any actionable decisions.
The simplest calculation involves dividing, over a given period, the number of transactions containing at least one item from the zone by the number of visitors who entered that zone. This remains useful as long as you keep its limitations in mind.
A single receipt often contains items from multiple zones and is therefore counted multiple times depending on the zones being analyzed. The total number of transactions per zone does not correspond to the number of receipts. Furthermore, a customer may purchase a product without stopping in the zone in a detectable way, or may linger for a long time without making a purchase. The resulting rate serves as a benchmark for comparison over time and between comparable zones, not an absolute measure of purchasing behavior.
Employees move through the zones all day without making purchases. If they are counted as visitors, the denominator increases and the conversion rate automatically decreases—all the more so when actual foot traffic is low. The effect is therefore greatest during off-peak hours—precisely the times when teams are trying to understand what isn’t working.
Several methods exist for excluding staff from the count without resorting to facial recognition, based on distinctive visual features, typical movement patterns from the stockrooms, or a combination of both. Regardless of the method chosen, it is validated through sampling before being implemented in production.
Four pitfalls recur regularly: attributing a rate variation to a promotional campaign when it is actually caused by a change in customer flow or camera framing; comparing two stores without adjusting for foot traffic and store format; Interpreting a rate from a time slot with low foot traffic as if it were stable. Changing the zone boundaries and then comparing them to previous periods, which is equivalent to changing the measurement instrument mid-measurement.
A simple rule limits the damage: any change to the zoning, framing, or counting method must be documented and accompanied by an explicit break in the data series.
The cross-referencing involves aggregated flows and transactions, not individuals. The video data used consists of counts and dwell times, not stored images, and no biometric processing is involved: facial recognition falls under Article 9 of the GDPR and is outside this framework.
Two points warrant attention. The cross-referencing must not allow an observed visit to be linked to a customer’s identity, which would be the case if the transaction were associated with a loyalty account and tied to an individual’s movement history. As for employees, the exclusion of staff from the count must not result in individual monitoring of employees: the Labor Code requires that employee representatives be informed and consulted regarding measures that may affect them.
CORE, XXII’s platform, analyzes footage from cameras already installed and generates aggregated metrics by zone, without facial recognition. These metrics—including counts, time spent in the store, and movement between zones—are intended to be cross-referenced with the store’s operational data, including point-of-sale data. The task of matching product SKUs to zones remains the responsibility of the retailer, as it depends on its store layout.
A useful initial cross-referencing process consists of four steps: synchronize the clocks of both systems; define five to eight zones corresponding to actual operational decisions; build the mapping table between product SKUs and zones; and then observe a period of at least four weeks before drawing conclusions. Discrepancies observed over a single week are most often due to noise.