Retail Establishments: How to Measure the “Store Visits” Generated by a Digital Marketing Campaign

Digital marketing campaigns on Google, Bing, and social media are not only useful for selling online. They are also very effective in attracting potential customers to commercial establishments.

This is highly relevant for the majority of businesses, as even though the growth of e-commerce is unstoppable, around 80% of retail sales still take place in physical stores, even for those transactions that begin online.

But if you want to carry out a campaign to attract customers to your establishments, make sure that you are going to be able to measure its impact with enough reliability and rigor. Otherwise, you won’t know if it is working or if you have wasted your money.

Online Sales in the US Retail sector. Percentage over Total Sales (source: U.S. Department of Commerce)

Online Sales in the European Union. Percentage over Total Sales (blue) and over Total Turnover (orange) (source: EUROSTAT)

Coupons: The Easiest Option?

A simple way to measure the store visits generated by your online advertising campaigns is with discount coupons. You can display promotions in your ads linked to a code that appears only in your campaigns. This way, you will know that anyone who comes to your establishment with that code comes from a digital advertising campaign. It is a method that is relatively easy to implement online, but it burdens the establishment’s workers and the accounting department with more administrative tasks, and it can create discomfort in regular customers who, waiting to be attended, see that others enjoy discounts or offers that they do not receive.

Automatic Measurement of Store Visits: THE SOLUTION?

Google offers a much better alternative in its campaigns: the automatic measurement of store visits. By utilizing the geolocation of users’ mobile phones and leveraging machine learning, Google can estimate with reasonable accuracy the number of customers who have visited your establishment after having viewed or interacted with your ads.

Its potential is immense. As a case in point, Google Ads has published on YouTube a case study on how Nissan has used this feature to attract customers to their dealerships and identify which campaigns and search terms worked best for achieving sales, by tracking their journey in the customer journey to reach them at the right time and in the right way. Results: a 25x ROI and a conversion rate on mobile phones (from ad click to dealership visit) of 6%. A staggering rate for the sector, given that the average buyer visits only two dealerships before buying a car.

But this solution is not a silver bullet, as the English speakers would say. That is, it doesn’t solve all our problems.

The first limitation is that it only applies to Google campaigns, leaving out campaigns that you conduct on Facebook, Instagram, Bing, Tik Tok, etc.

The second problem is that, for Google to share physical store visit data with you, you must meet a series of requirements, among which the following stand out:

Your stores must be located in one of the countries where this feature is available (as of today, there are 27 countries, including Spain and 12 other European countries, but only two in Latin America: Chile and Mexico).
They must not be of a nature considered sensitive. For example, stores where products related to healthcare, religion, sexual content, and childhood are sold are considered sensitive and, therefore, do not meet the requirements for store visit conversions.
Typically, Google only offers this functionality to clients with more than one establishment in the country. The help from Google Analytics is explicit on this point: “Have multiple establishments located in qualifying countries.” It is not clear that this remains the case for Google Ads campaigns, but what the Google Ads help says as of today is that “the more stores you have, the more likely your account will have enough data to use this feature.”
But probably the most limiting factor for many advertisers is the required traffic volume. Google Ads says the following: “To overcome our privacy thresholds, your ads must have enough clicks or impressions, and your business must have enough foot traffic. A machine learning model is used to register store visits, so we need a sufficient volume of data to accurately record this type of visit. This figure varies by advertiser, as each one is different.” As you can see, everything is quite vague. Again, the help from Google Analytics is more explicit, indicating that “generally, at least 100,000 website sessions in 30 days are required for store visits to appear in the reports”. On the other hand, our experience with some of our clients is that, without reaching this threshold, after 6-7 months generating between 20 and 40 thousand visits from Google Ads (and a total of 50-70 thousand sessions including organic and direct traffic), Google Ads began to share store visit data.

In summary: accessing this data is not immediate or free, as it requires several months of significant investment in Google Ads campaigns, nor is it available to everyone, as there are significant geographical and sectorial limitations. It is also not an option for businesses whose potential clientele is relatively small, as they will never reach the necessary traffic threshold.

You may spend months seeing this message in the Google Ads interface while waiting for the physical store visit measurement to be activated

Some Simple and Frequently Used (but WRONG) Alternatives

Pre-Post Evaluation

Throughout history, we have seen in the press the assessment of the goodness of certain political or sports decisions by comparing the before and after. For example, the regularization of electricity prices or the dismissal of a football coach. If your team scores more points with the new coach, it is assumed that the change has been a success. In the world of impact assessments, this is called a Pre-Post evaluation. It involves measuring the differences in the outcomes of a program by comparing the situation before and after its implementation.

The problem with this method is that it assumes that there are no other external factors affecting the outcomes beyond your intervention. For example, if your team improves its performance after changing the coach, it might be due to the recovery of previously injured players or the fact that they faced easier opponents. In such cases, you might be giving the new coach credit that isn’t entirely deserved.

In the world of advertising, if you sell umbrellas, you launch a digital marketing campaign, and it starts raining for a month, did you increase sales because of the campaign or because of the rain? Or was it due to word-of-mouth marketing? In a recent real-life case with another one of our clients (a restaurant chain in Madrid), the data for one of their locations in July (where we were running a local Google campaign) showed zero growth compared to the previous month. A pre-post evaluation would have erroneously concluded that the campaign had no impact. The reality was that in Madrid, in July 2022, during a heatwave, the streets were empty, and restaurants were deserted at lunchtime. The campaign actually prevented an estimated 25% drop in sales, which was evident in other “control” restaurants where we didn’t run any advertising.

Simple Difference

Another common method of impact evaluation in everyday life is that of Simple Difference. It involves comparing results between two groups (or two locations); one of which receives the program, and the other does not.

The Simple Difference method implicitly assumes that the two groups are identical and subjected to identical external conditions, except for the program whose impact is being evaluated. This is almost never the case, except in randomized controlled trials (known as RCTs in their English acronym).

During the pandemic, we have seen countless of these “poorly done” comparisons: if Sweden did not implement a lockdown while Spain did, and there are more deaths in Spain, it means the lockdown didn’t work. If Rio de Janeiro had no lockdown while Madrid did, and there were more deaths in Rio, it means the lockdown worked. And so, the media exchanged examples and counterexamples ad infinitum. But we are neither Swedes nor Brazilians, and apart from the lockdown measures, there were plenty of geographic, demographic, sociocultural, and public health factors that invalidated the comparisons, starting with the level of pandemic penetration in each place at the time of implementing measures, population density, socialization habits, intergenerational cohabitation in households, climate, etc.

In our industry, the owner of multiple commercial establishments might be tempted to conduct a pilot test of a digital marketing campaign to attract customers to one of their stores and compare the results with their other establishments. However, without controlling for socioeconomic, demographic, cultural, and/or geographic factors, the estimation will be erroneous. For example, a dealership conducting a campaign for an electric vehicle model in an urban and high-income location would overestimate the campaign’s impact if they compare the results achieved with sales from another dealership located in a rural and economically depressed area where no campaign was conducted.

This example may be very obvious, but we continuously encounter clients requesting A/B tests with divisions like North-South or Center-Periphery in Spain, as if there were no substantial underlying differences between the various regions.

Much Better Alternatives

Differences in Differences

A more robust and rigorous evaluation method, compared to what Pre-Post or Simple Difference offer but still affordable and straightforward to implement, is the method of Differences in Differences (DD or DID). In a way, it combines the two previous methods because, like in Simple Difference, it starts with two groups (treatment and control), and, like in Pre-Post, it measures before and after the intervention. However, in this case, it also measures before and after in the control group, where no action has been taken. This way, it allows for controlling external effects that may have affected both groups simultaneously. For example, in the case of an advertising campaign conducted in certain establishments, it estimates the comparison of sales trends in these (treatment group) with the trends in those where no advertising was done (control group).

Illustration of the Differences in Differences Method

This method takes into account that there may be initial differences between the two groups and external factors that affect sales independently of advertising campaigns, but it assumes that such factors will equally affect both groups. That is, in the absence of advertising, both groups would follow parallel trajectories. This is usually a reasonable assumption when the context of both groups is sufficiently similar, and the duration of the campaigns is not very long.

Synthetic Control

However, the aforementioned assumptions are not always met. For example, in the case of sudden economic crises like those experienced in recent years, within the same geographic area, some sectors may be more severely affected than others, leading to significant divergence in the trends of the treatment and control groups. In such cases, the Differences in Differences method ceases to provide valid results, and we need to resort to more complex methods like the Synthetic Control, a methodology developed at MIT by the Basque academic and researcher Alberto Abadie for impact evaluation studies in the social sciences, and adapted for digital marketing by Fáktica Analytics.

Synthetic control allows estimating the impact in situations where a single unit (a country, a city, or a establishment) benefits from the intervention (in our case, an advertising campaign). Instead of comparing this unit to a group of units that have not received the intervention, this method uses the most relevant characteristics of the treated unit and the untreated units to construct a “synthetic” or artificial control unit, assigning relative weights to each of the untreated units in such a way that the synthetic control unit closely resembles the treated unit. This combination of untreated units that recreates a synthetic control results in a better counterfactual than each of the individual real units by themselves.

How do we do it at Fáktica Analytics?

Depending on the circumstances of each case and the robustness of the hypotheses, we apply either the Differences in Differences (DD) method or the Synthetic Control method. For example, we used the DD method to measure the impact of online advertising on the mentioned restaurant chain in Madrid because climatic and seasonal effects affect both groups similarly. In this case, we adapted it to not only control for seasonal effects but also for the word-of-mouth effect, which grows at a higher rate in recently opened establishments than in those that have been in operation for years.

This way, we can recreate the counterfactual, that is, answer the question: What would have happened in the establishment we advertised if we hadn’t run the advertising?

Illustration of the Counterfactual (real case)

Solid red line: number of reservations in the advertised establishment (actual data). Dotted red line: counterfactual estimate (expected reservations without advertising, accounting for seasonal effects and word-of-mouth). The impact of the campaign is the difference between the two lines.

This analysis was crucial in helping our client calculate the return on campaigns, decide on their continuation, and determine the budget to invest in them. A proper measurement of campaign impact is critical for making the right decisions. Without it, we might be wasting money on campaigns that have no positive impact, or conversely, we might be leaving money on the table, missing out on growth opportunities by stopping (or not scaling) advertising campaigns with more impact than we initially suspected.

If you want to know how we would apply it to your establishments, request an analysis from us. You can contact us without obligation. We will be happy to assist you.

——–

About the author of the post:

Marcos Ferreiro is the CTO of Fáktica Analytics and a professor of Program Evaluation and Impact Measurement at IE University.