Since we launched our Facebook Political Ad Collector project in fall 2017, more than 16,000 people have participated in it. They all agreed to install a browser plug-in that anonymously sent us the ads they see when they browse Facebook. We used that data to understand and report on how political messaging on Facebook works, and how the system is being gamed to manipulate the public discourse.
Although the number of users is large, over the summer we noticed a potential snag: We were receiving more ads from Democrats and progressive groups than from Republicans or conservative groups. Our hunch was that this was because we had more liberal participants than conservatives ones.
We tried a number of things to make our ad collection more diverse: to start, we bought our own Facebook ads asking people across a range of states to install the ad collector. We also teamed up with Mozilla, maker of the Firefox web browser, for a special election-oriented project, in an attempt to reach a broader swath of users.
But because we made the Political Ad Collector almost completely anonymous, we couldn’t say much definitively about the audience. We also couldn’t know for sure how much of the skew in our findings was because of the people who participated in the project and how much of it was because left-leaning groups used Facebook advertising more than (or differently than) conservative groups.
To dig deeper, we partnered with some academic researchers and a research firm called YouGov to create a panel of users from across a wide spectrum of demographic groups and political ideologies who would agree to use a special version of the ad collector plug-in that was less anonymous. For these users, we had a unique ID that was tied back to data about them provided to us by YouGov — demographics, political leaning, race, state of residence, but not name or address.
This collaboration was suggested, and funded, by the Democracy Fund.
YouGov paid members of this new group to install a special version of the Political Ad Collector browser plug-in. We were able to link their answers to demographic questions — like age and partisanship — to the ads that they sent us. Through this survey, 3,588 participants submitted at least five ads from Oct. 15 until the election. The group was designed to yield a large and diverse enough sample that will let us, and academic researchers who study political ads, make statistically valid claims about the ads people see.
We didn’t change anything about the way we collect data from users of the original, publicly available ad collector plug-in who were not participating in the YouGov survey. They remain anonymous to us, and we collected absolutely no personal information from them.
Ads that were seen by participants in the YouGov survey, with the demographic data stripped, are part of our existing database of ads. We aren’t publishing the survey-derived data, though we have shared it with some academic researchers.
Now that we’ve got a better sample that can help us draw broad conclusions about Facebook political ads, here’s some of what we’ve learned from the project:
Perhaps unsurprisingly, more than 70 percent of all the political ads we saw were highly targeted by ideology. Most ads were shown to at least twice as many people from one side of the political spectrum than the other. Only about 18 percent of political ads were seen by anything close to an even ratio of liberals and conservatives.
One advertiser that targeted both sides was AARP, which spent about $700,000 on ads from May to the election. Many of those ads simply urged viewers to vote; some ads encouraged people to hold their member of Congress accountable for voting yes on “last year’s bad health care bills.” The AARP has opposed efforts to repeal and replace the Affordable Care Act.
“AARP has a long history of nonpartisan voter engagement,” an AARP spokesperson said in a statement. “Our goal was to engage our members, whose views span the ideological spectrum. These facts guided our strategy.”
The ad that was seen by the most people in the YouGov sample was this one from Tom Steyer’s “Need to Impeach” organization, which included a video saying, “We need to impeach Donald Trump before he does more damage,” citing “legally segregated camps” for migrant children and Hurricane Maria deaths. The ad was seen almost exclusively by self-identified liberals in our sample. According to Facebook, Steyer has spent $2.3 million on Facebook ads between May and Election Day. (Steyer and his wife, Kat Taylor, are donors to ProPublica, and Taylor is on our board.)
But because the Facebook advertising system makes it easy to show ads to very small target audiences, a relatively small number of participants in our project saw the “Need to Impeach” ad. Of the 3,588 people who submitted at least five ads from Oct. 15 until Election Day, only 146 saw it.
That microtargeting may help explain why, even with a pretty large number of participants, and including all these additional participants who better represent the country, we still didn’t see any ads or advertisers caught up in investigations and news stories about foreign election meddling. We did, however, report on a mysterious Facebook page called “America Progress Now” urging liberals to vote for Green Party candidates. The candidates themselves had never heard of the group, and we couldn’t find any address or legal registration for it.
We also saw ads from liberal groups that used misleading tactics we first saw being employed in 2016 by groups like the Internet Research Agency in Russia. “Voter Awareness Project” urged conservatives not to vote to re-elect Ted Cruz, the Texas Republican senator, citing Trump’s previous antagonism toward him. But the group was actually run by a prominent liberal. And other liberals, like Ohio gubernatorial candidate Rich Cordray and the Environmental Defense Action Fund, ran political ads from pages with names that implied they were news organizations, like “Ohio Newswire” and “Breaking News Texas.”
We also found a fair number of political ads that Facebook didn’t label as political as part of their new system, including clearly political ones from Sen. Kamala Harris, Uber, Alliance Defending Freedom and others.
We still have some unanswered questions about how advertising works on Facebook, including some that go beyond political ads:
While we found a way to determine, in part, how an ad is targeted, there are complexities to Facebook’s systems that we can’t detect or understand. For instance, what is the impact of the algorithm Facebook uses to show ads to whoever is mostly likely to click? Upturn, a think tank that researches equity issues in the design and governance of technology, says in a court filing that they ran a job ad for a bus driver and, on its own, Facebook showed this ad to four times as many men as women.
Advertisers pay Facebook more for their ads to be seen by some kinds of people than by others — based on, for example, age, demographics and median income. What’s the effect of some people seeing on average more expensive ads than others? How big is the disparity? Whose attention is cheap, and whose is expensive? How do cheap ads differ from more expensive ones?
Facebook has a “lookalike audience” ad-targeting feature, which lets advertisers target people who are “similar to” their page’s visitors or personal contact info they upload. Could this be a sneaky (or even unintentional) way to target a housing or job ad in a discriminatory way?
Facebook classifies individual users by their political beliefs and their “multicultural affinity,” letting advertisers show ads just to, for instance, people who are “very conservative” or who have “Multicultural affinity: African American (US).” How accurate are Facebook’s guesses? You can see political ads targeted by multicultural affinity here: African-American, Asian-American, Hispanic - All, Hispanic - English Dominant, Hispanic - Spanish Dominant, Hispanic - Bilingual.
One final note: Now that the midterm elections have finished, with only a few U.S. elections happening in 2019, we’re busy figuring out the Ad Collector’s next chapter. We’ll make an announcement shortly.