The Evolution of Location-Based Targeting
More and more marketers and developers are using mobile device coordinate data to better understand their customers or users based on how they move through the physical world. This not only includes visits to retail stores, but shows a complete picture of real world movement, including when a device is at home, at work, at a park, at a school, traveling for business or fun, and yes, at a Starbucks or Taco Bell.
Mobile device coordinate data is proving to be an incredibly valuable signal, and with the continued growth of smart devices (not limited to phones) and the time we all spend on them, the trend shows no sign of slowing. At the same time, the technical methodologies and approaches for answering the question of where a device is (or what the user is doing) are not all created equal, and it’s worth taking a look at the different methodologies and how they are evolving.
Location Targeting Started with Geofencing
When marketers think about location-based targeting, they often think of geofencing. The term has become a catchall for a basic location identification tactic, but it is much more nuanced than that. A geofence is really a virtual geographic boundary, typically defined either as a point and radius (meaning a circle) or as a polygon (meaning a complex shape). Geofencing services take a device’s estimated coordinate data (latitude, longitude), which can come from the Global Positioning System (GPS), be triangulated from cell towers, or be interpreted from a wi-fi hot spot or beacon signal (all of these with various trade-offs between scale and accuracy), and calculates whether the device is inside a defined area. If it’s a match, advertisers make an optimistic bet that the user is at a “place” — a point of interest, business, et cetera.
Hello Machine Learning
Machine learning has been driving improvements in nearly every aspect of technology — and location-based targeting is no exception. Machine learned place attachment is a more sophisticated approach to understanding where a device actually is in the physical world and warrants some time exploring. Let’s take a closer look at place attachment: how it works and how it differs from geofencing.
Machine learned place attachment is based on training models that take into consideration data from both the mobile device and the places around it. Models are trained on annotated data, meaning data from confirmed users at specific places and their corresponding device data and place data. Examples of data from the device include GPS coordinates; time of day; accelerometer data; Wi-Fi access point names and signal strength; beacon signals, barometer data and motion detection (e.g., walking, biking or driving). Examples of data about the places themselves include their coordinates, category, popularity at various times, operating hours, polygon or parcel tracing, and neighboring physical structures (rivers, roads).
Well-trained models learn nuances like the following:
- Devices that are driving are not at a certain class of places.
- Movie theatres are more likely to be attached at night.
- Schools are more likely to be attached during the day.
- Polygons are more useful for certain categories of places.
- GPS signals are less useful in certain urban densities, and should be based on place types, countries, etc.
By taking a machine learning approach, you wind up with a scalable and continuously improving process that is more accurate than simply creating a few rules, which in turn helps marketers craft more effective messaging and better understand their audience.
Apples versus Oranges
Don’t get me wrong, geofencing is very effective for certain use cases that don’t require assigning a device or person to an exact place, i.e., if you want to determine if devices are close enough to drive to a store, but it should not be considered the best method for understanding if a user is actually at a given location.
The distinction between geofences and place attachment is reminiscent to the early evolution of search when we went from directories to real search engines. Organizing the internet (and I’m dating myself) began by humans manually curating directories. Guess what? … it didn’t scale and wasn’t accurate. Quickly, Google, AltaVista, Inktomi and others developed machine learned models to determine what website was most relevant to a query. Similarly, we see machine learning based place attachment quickly replacing rules based geofencing to better understand a popular marketing query: Where is my customer?
Geofencing isn’t the only approach to location-based marketing, nor is it the most advanced. Machine learned place attachment is proving to be incredibly effective, and should be added to the marketing lexicon as a methodology for understanding your customer’s real-world behavior. Better data makes for better customer experiences.
Bill Michels is SVP of product & partnerships at Factual.