Understanding How Google Measures Store Visits
At the LSA Place conference in New York last week, Google’s Kishore Kanakamedala, director of online-to-offline solutions, offered a rare glimpse into the procedure Google uses to link online ads to offline store visits.
Since the launch of store visit metrics in late 2014, Google has occasionally pulled back the covers to discuss how attribution works, and the presentation was a welcome update, touching on the steady growth of the dataset and the addition of new techniques such as machine learning.
Google has collected data on more than 7 billion store visits in 19 countries, up from 4 billion as reported in March. Even though ecommerce is growing and brick and mortar retail is arguably in the midst of a slow decline, 90% of consumer dollars are still spent in physical stores, and the intent of Google’s store visits data is to help demonstrate the efficacy of multiple online touchpoints that might drive consumers from their laptop, tablet, or phone into a store.
With 3 in 4 mobile searchers visiting a store within 24 hours of searching, Google recognizes that consumers who search locally have a high likelihood of converting. Expanding upon this, the company believes that any incentive with a local hook should be measured by its ability to bring customers into stores.
Google places great value in consumers who view local ads, claiming that such consumers are 40% more valuable overall than people who just walk into a store without viewing an ad, and 25% more likely to make a purchase in the store. Attribution measures store visits linked to search ads, Maps ads, display advertising, Google Shopping results, and as of a couple of weeks ago, local ads on YouTube.
Attribution metrics can help measure the return on investment for advertising campaigns run by local stores or multilocation brands, and can also be used by manufacturers, who can employ affiliate location extensions to track visits to stores that carry the manufacturer’s products.
Using the same technology that underlies Google Maps, Google has measured the coordinates and geometries of “a couple hundred million buildings in the world,” according to Kanakamedala. The combination of Google Earth data, Street View data, and in-store Wi-Fi scans allows Google to pinpoint whether a user is in a store or outside right next to it. Google now understands the borders between stores with a high degree of confidence.
This confidence is in large part due to what Kanakamedala referred to as Google’s “upgrade from shallow learning to deep learning,” which, he said, has given the attribution algorithm the ability to train on very large datasets. Google ingests and correlates a wide combination of data points to derive attribution, including GPS signals, cell tower signals, Wi-Fi strength signals, location history, visit duration data, search query data, and Bluetooth beacon signals.
Google uses machine learning to reduce noise in the data. For example, a user who passes through a store on the way to the parking lot and only spends one minute inside should not be counted as a visitor, and neither should an employee who spends hours there. Kanakamedala suggested that Google has gotten better over time at detecting such outlier conditions and training the system to recognize the distinguishing characteristics of legitimate visits.
Google says that all of this visit data is fully anonymized, with no data on individual users making its way into the aggregate visit metrics. Regardless, users who don’t want to be included in attribution metrics can opt out by turning off location sharing.
To confirm the findings of its primary data sources, Google makes use of a panel of more than 1 million users who have opted to share their full location history. Ground-truth data from these users helps to verify and improve the results derived from other signals.
With the expansion to local YouTube ads, attribution now covers a broad range of advertising touchpoints. Consumers may, in fact, be encountering a brand in a variety of ways before a visit occurs, and Google has begun measuring the amplifying effect of local ads viewed in succession.
As valuable as attribution data can be for advertising, I find the broader applications of the technology to be of even greater interest in the long term. Advertisements are only one of the many properties in the Google ecosystem that could be correlated with visits to physical locations. The applications for this data will undoubtedly continue to grow.