A method note from SnowBright.
The problem with most in-store analytics
Most reporting on physical retail describes what already happened and calls it insight. Traffic was up. Engagement was strong. The campaign performed well. None of these can be acted on, because none of them answers the only question an operator actually has: compared to what?
A number without a comparison is a description. A number against a baseline and a control is a measurement. Everything below is about that difference.
What we measure
We report a small set of metrics, defined the same way every time:
- Footfall — counted entries to a defined space, aggregated. Not unique persons.
- Zone dwell time — how long people remain in a defined zone, reported as a distribution rather than an average, because dwell is rarely normally distributed and the median tells a different story from the mean.
- Opportunity-to-see (OTS) — the count of people passing within a defined viewing geometry of a display, following the conventions used in out-of-home measurement.
- Attention seconds — dwell within that viewing geometry, segmented by creative and by day-part.
- Zone sequencing — the order in which zones are visited, which is what makes path-to-purchase an observation rather than an assumption.
- Conversion — transactions over footfall for the same window, taken from the point-of-sale system rather than estimated.
Definitions are fixed before a study begins and do not change mid-flight. A metric redefined during a campaign cannot be compared across it.
The measurement design
Baseline. Before any intervention we record traffic by day-part, dwell distribution by zone, and conversion by category, for long enough to capture weekly seasonality.
Control. Treatment locations are matched to control locations on footfall, format and catchment. Where a matched control is impossible we use the site’s own pre-period, and say plainly that the design is weaker for it.
Window. The measurement window and the primary success metric are fixed in writing before the campaign runs. Choosing the metric after seeing the data is how organisations convince themselves of results that are not there.
Effect. We report the difference between treatment and control with a confidence interval. Where the sample cannot support a conclusion, that is the finding.
Decision. Every report closes with the action it supports: extend, revise, relocate or stop.
What behavioural science adds
Measurement tells you what happened. Behavioural science is how you generate the next thing worth testing.
The interventions we help design draw on a small, well-evidenced set of mechanisms: salience, what captures attention in a cluttered field; anchoring, how a first price frames every price after it; social proof, how the visible behaviour of others changes behaviour; choice architecture, how the ordering and defaulting of options shifts selection; and friction, how small costs at a decision point suppress action out of all proportion to their size.
These are hypotheses, not guarantees. Effects that replicate in a laboratory frequently shrink or disappear in a shop. That is exactly why each one is run as a test rather than installed as a certainty.
What we do not claim
- We do not attribute a sales change to a display without a control.
- We do not report averages where the distribution is skewed.
- We do not identify individuals, and our measurement is designed to work without it.
- We do not present a result as significant when the sample cannot carry it.
Measurement that only ever produces good news is not measurement. It is marketing with a chart on it.
Privacy
Our measurement is built to work without knowing who anyone is. We do not use facial recognition and we do not collect personal identifiers. Audience data is aggregated and anonymized at the point of capture.