I have often been troubled by the lack of a clear definition of political micro-targeting. There is a lot of discussion of the macro effects of micro-targeting, and of the larger impact of data-driven elections. It arguably serves to fragment political discourse, to accentuate wedge issues, to promote filter bubbles and leads to a transactional politics where localized claims and promises remain unchallenged.
Much of the literature tends to assume that political micro-targeting requires a precisely segmented audience, a specific location, and, most importantly a focussed policy message. If that is the case, then how much political micro-targeting actually occurs in Canadian elections? And in particular, what do these ads look like to the average Canadian voter?
With the excellent assistance of one of my graduate students (Jesse Gordon), we examined a sample of ads from the public Facebook ad library posted during the 2019 federal election. A longer paper reporting the results of the research has now been posted on my SSRN page.
Rather than creating a developer account, or installing a browser plug-in, techniques used by the Digital Democracy project in their report on Understanding the Digital Ecosystem, we used the public version of the ad archive. This version provides a grid-style list of advertisements that can be queried based on a keyword or page search. Additionally, the ad archive filters search results by country, whether the ad is active or inactive, the number of impressions, whether the ad had a political disclaimer, and the platform on which the advertisement was displayed. The advertisement image and message are displayed alongside additional information such as the ad ID, date range of when the ad was active, who paid for the advertisement, how much was spent on it, and crucially for our purposes, the provinces in which the ad was displayed, and the age range and gender of the people who saw this ad.
Facebook also reports the number of impressions for each ad. Facebook defines impressions as the total number of times the ad referenced has been shown on the site. Thus, when an ad appears on the right side of the screen while a user is viewing a Facebook product, an impression is registered to that Facebook account. This metric is distinct from the number of times the ad is clicked on, which is not information provided by the library.
We selected two dates for each major political party: the first large ad push (the first day in the election period, in which the political parties posted the largest number of individual advertisements), and the last large ad push for each party: September 10 and October 18 for the NDP resulting in a sample of 45 ads; September 7 and October 19 for the Liberal Party resulting in a sample of 185 ads; and September 6 and October 20 for the Conservative Party resulting in a sample of 77 ads. For the Green party, because of their limited Facebook ad spend, every ad was collected resulting in a sample of 35 ads. We also examined ads in eight separate ridings to examine whether the patterns of digital advertising differed depending on whether the riding was perceived as “safe” or “swing.” Our research yields three general conclusions.
First, if political micro-targeting does entail a relatively focussed demographic audience, a focussed location, and (critically) a precise policy of issue claim, then we found that only a small minority (7%) met these criteria. Most ads critically lacked precision on one, or two, of these critical variables. If a general appeal is targeted at a particular location and/or a particular audience (as many were), should such an ad be classified as political micro-targeting? Further, if a more focussed policy message is delivered at a broader demographic, in more widespread locations, is that political-microtargeting?
These questions are relevant because one of the major risks of micro-targeting is that candidates will make one claim or promise in one place, and another elsewhere. This, it is feared, would encourage a more transactional politics, where promises are made in precise locations to precise groups of swing voters, with little regard for national policy messages. Political micro-targeting allegedly fragments political discourse, eroding the public and national conversation about the public interest. According to Susan Delacourt, it encourages shopping for votes. And yet, we found virtually no example of an ad that promised a precise and localized reward if one candidate were elected. That kind of transactional micro-targeting was largely not present in the 2019 Canadian federal election — at least it was not apparent on Facebook. To the extent that claims about policies or issues were made, they were generally pitched at a relatively abstract, and largely uninteresting, level with little distinction between the messaging on Facebook and that delivered through other media.
Secondly, this finding supports the conclusion that the resources for precise content-creation that might be precisely targeted are not available to the main Canadian parties. It is one thing to segment a likely audience effectively, but another to create the targeted issue-specific content that might be directed to such an audience. This limitation is a significant one, even in the world of resource-rich candidates and parties in the U.S. The constraints are even more severe in Canada.
The third conclusion relates to the reporting of the ad impressions. Our sample showed a range from less than 1000 to nearly 1 million ad impressions. It is tempting to conclude that a smaller number of impressions is related to a more focussed audience sample and a higher degree of micro-targeting. But this would be misleading. In fact, to the outside viewer, there is no way of knowing whether a small number of impressions is an indicator of a focused audience segment.
There is a confusion, therefore, between the obscure micro-targeting processes used by Facebook and the visible manifestation of the ads to the average voter. There is also confusion about what in fact we mean by an “ad”: this complexity complicates any methodology for counting and comparing “ads.” A Facebook ad is less a discrete message and more a complex machine for producing further messaging. Modern political advertising certainly begins with a list of potential target users (the “custom audiences”), but that is just the starting point of a complex, murky and continuous process through which the Facebook algorithms learn and deliver subtly different messages dependent on a multitude of different variables.
The Facebook political ad library provides access to unprecedented levels of information about political digital advertising in Canada. It has clearly has had some positive effects. We did see fewer examples of dark posts. Obvious examples of disinformation were exposed. However, the archive provides little practical guidance to the voter about why I am seeing this ad. Despite these efforts at transparency, we are only marginally more informed about the extent and nature of political micro-targeting occurring in Canadian elections. These findings mirror research in other countries, and support calls for further efforts at ad transparency in the years ahead — such as those made by Privacy International.
Our findings suggest not only, however, the need for higher levels of transparency, but also a more nuanced understanding of how the machinery of ad generation works, especially in countries outside the U.S. Not all micro-targeting carries the same precision. And not all raise the same concerns about electoral manipulation and propaganda.
Critically, our analysis also suggests a higher level of transparency, not only for the social media platforms, but also for the data analytics performed through the political parties’ voter relationship management systems. We know very little about the interaction between the data analysis performed through databases like Liberalist, for instance, and the customization of “Look-Alike” audiences on Facebook. Greater visibility about “why I am seeing this ad” requires an understanding of how the parties themselves generate the assumptions about the kinds of people who might be persuadable through digital advertising. Our research supports the wider calls for the inclusion of federal political parties into the wider framework of privacy regulation in Canada.