Measuring a store visit you cannot see
Most purchases in the region still happen in a physical store. Any measurement model that stops at the click is measuring the minority of the outcome.
E-commerce penetration across Latin America has grown quickly and is still, in most categories, a minority of total retail. For groceries, pharmacy, home improvement and a long list of others, the overwhelming majority of transactions happen in a building.
Which means that a campaign measured entirely on click-through and online conversion is reporting on the small end of what it did.
What people usually do instead
Faced with this, most advertisers pick one of three coping strategies.
Measure online and assume the rest. Report the digital conversions, describe the offline effect qualitatively, and hope nobody asks for a number. Common, comfortable, and it caps how much budget can ever be justified.
Use a proxy. Store locator clicks, “check availability” interactions, coupon downloads. These correlate with visits, sometimes well. They are also easy to inflate accidentally, and everybody in the chain knows it.
Run a matched-market test. Genuinely rigorous, genuinely slow, and it needs enough comparable markets to be statistically meaningful. Excellent once or twice a year. Useless for optimising a campaign in flight.
Each is defensible. None of them gives you a number you can put next to online conversions in the same report.
Panel-based visit measurement, honestly described
The approach that scales is inferring visits from location panels: a set of users whose device location is available with consent, matched against store polygons, compared between an exposed group and a control group.
It works, and it is worth being precise about the conditions under which it works.
It is a sample, not a census. The output is a modelled lift with a confidence interval, extrapolated from panel behaviour. Anyone presenting it as a headcount is overselling it.
Panel composition matters enormously. Location panels skew toward certain devices, certain app categories and certain income bands. In markets with a wide device spread — most of ours — this is a real source of bias that has to be corrected for and disclosed.
Store polygons are harder than they sound. A supermarket inside a shopping centre, a pharmacy on a busy corner, a store with a shared car park. Poorly drawn polygons produce confident nonsense.
Control groups are not optional. Without a holdout, you are measuring the seasonality of the category and calling it campaign performance. This is the single most common failure, and it always flatters the campaign.
What a usable read looks like
A store-visit measurement worth acting on has, at minimum:
- A pre-registered control group, defined before the campaign runs rather than assembled afterwards.
- A stated confidence interval, and a willingness to report that a result was not significant.
- Panel composition disclosed against the market’s actual device and income distribution.
- Visit definitions written down: dwell time threshold, polygon method, exclusion rules for staff and delivery traffic.
- A separate read on incrementality, because a visit that would have happened anyway is not a result.
The reason to bother
Precision is not really the point. The point is proportionality.
If online conversion is fifteen percent of the outcome and it is the only fifteen percent being measured, the campaign is being judged on a fraction of what it did — and the budget will be sized accordingly, forever.
A modelled, honestly-bounded read on the other eighty-five percent is more useful than a perfectly precise read on the wrong slice. Advertisers who accept that trade tend to fund larger, longer programmes, because for the first time the reporting is describing something close to the whole business.