Facebook users in five key swing states have been seeing a peculiar sequence of political ads pixelating their news feeds for the past six months.
It begins with a carousel ad from a page called United Research Group asking them to fill out a lengthy survey. Soon afterward, multiple ads from Pacronym, a progressive super PAC, begin to litter the Facebook experience of about half of those who had been surveyed. Then, an ad for a different but related survey appears.
This rather specific experience is an intentional and coordinated effort to reach persuadable voters in critical presidential battlegrounds, a result of months of work by a group of former Facebook employees and data scientists at the progressive nonprofit group Acronym.
Essentially, the group, which includes some who worked on the Trump campaign in 2016, has co-opted the political ad function on Facebook to perform real-time persuasion message testing, to get a sense of how voters are reacting to ads as they see them.
In the fast-paced world of digital advertising, the availability of real-time data beyond mere engagement is fairly small, leaving campaigns with a patchwork of clicks, old polling and hunches to assess the impact of the millions of dollars they are spending on digital platforms.
And Facebook, with about 220 million users in the country, remains the central digital vehicle for reaching Americans who are spending more time online during the coronavirus pandemic. About $100 million has already been spent on the platform for the 2020 presidential election.
The coronavirus outbreak has also forced 2020 campaigns to rely on a nearly entirely digital infrastructure, from fund-raising to organizing to persuasion. Having fresh data to inform campaign arguments online is essential.
This real-time testing project aims to help fill that gap. Called “Barometer,” the project has been the obsession of James Barnes, the former Facebook employee who was heralded as an “M.V.P.” of the 2016 Trump campaign and has since dedicated his professional life to undoing the results of the last presidential election. He found a home for the project at Acronym, which also houses Pacronym.
“The thing that Facebook does really well, and the thing that I learned pretty extensively at Facebook, is how to measure things,” Mr. Barnes said. “As I considered what I wanted to do after Facebook as I went through sort of this political transformation, I wanted to know: What is the best way that I could contribute to help defeat Trump in 2020?”
During the 2016 election, Mr. Barnes often used a Facebook tool known as “brand lift” to test different persuasion messages online for the Trump campaign. The tool provided insight into whether some of the online ads were moving people.
That tool has since been taken away from political campaigns by Facebook, part of its broad restructuring of how campaigns are allowed to operate on the platform following widespread backlash after the 2016 election.
Through his small team of engineers and data scientists, as well as ample cash from Acronym, which does not disclose its donors, Mr. Barnes has been able to recreate, to an extent, a similar tool.
“What’s felt impossible in the past is to test the real-world effect of ads of people who actually saw them, on an ongoing, consistent basis,” said Solomon Messing, the chief data scientist at Acronym. “And what we did was to develop a way to do this on Facebook.”
The tests are run in five states — Arizona, Michigan, Wisconsin, Pennsylvania and North Carolina (a full test involves selecting the audience, delivering an initial survey, running ads to half of the respondents, running the second survey to everyone and analyzing the results).
Using a voter file maintained by Acronym, the group casts a wide net for a list of names, based in part on voting history, propensity to vote for Mr. Trump in 2020 and party registration. The group then sends those results in a list to Facebook to build an audience to advertise to in those five states.
The audience is further whittled through the survey, which features fact-based questions like who controls the House or the length of a Senate term. The audience that is unable to answer those questions is considered among the most persuadable with ads.
The findings show that voters with low political knowledge and low news consumption, Mr. Barnes said, “support the idea that when folks have just kind of a lower baseline of information, we actually can persuade them by just showing them more of the information that’s true.”
Of course, these tests are run in relatively small groups, and the overall impact of the ad campaigns that result from these findings would be nearly impossible to assess at scale. And low-information voters, one of the central targets of the program, can be notoriously fickle in their preferences, particularly early in the year.
Additionally, the Barometer team knows its testing involves a lot of “noise,” the statistical term for unexplained variables that could muddy their results.
But seeing overall trends and percentage points of moves, they say, is a similar process to bringing a prescription drug to market: a test group, a control group and then results, with conclusions drawn from there.
Barometer has run a variety of tests on messaging during real-time news events, from impeachment to the killing of Qassim Suleimani, the Iranian military commander, to the coronavirus response. The results haven’t always been as expected.
Nearly all messaging about impeachment was received poorly, and low-information and swing voters tended to side with the president. Pacronym quickly dialed back some of its impeachment advertising.
The most consequential test, however, was over the killing of Mr. Suleimani. Though responses mostly hewed to partisan lines, the team saw significant movement among Trump-leaning voters away from the president when presented with critical commentary from a conservative messenger. It is a tactic known as “boosted news,” or the practice of paying to place news articles in the newsfeed of users.
In this test, a clip featuring Tucker Carlson, the prime-time Fox News host, excoriating the president for coming close to starting a war with Iran lowered the president’s approval rating among older voters and low-information voters in their test universe. Trusted local news sources also had an especially big impact.
More tests, particularly in the coronavirus response, found that boosted news had a far greater impact than running a typical negative political ad with all the hyperbolic trappings. That result excited, among others, the accountants at Acronym; simply boosting news and conservative commentary critical of the president was much cheaper than creating original ads and content.
Though most of the results have remained in-house at Acronym, and informed Pacronym’s spending, the group has started briefing other Democratic strategists about the findings, and is hoping to make them public soon.
“There is this sort of need to hold any research close to the chest that I think is a bit outdated and irrelevant today because what works for our voters to get them to vote for Vice President Biden in November is not going to work for Trump’s purposes,” said Tara McGowan, the chief executive and founder of Acronym.
Democrats hope the Barometer tool can help them improve a major deficit they face online against Republicans, what’s known as the “last mile” of content to reach a voter. Whereas a Trump tweet ricochets around a vast, expansive and influential far-right echo chamber online, Democrats have been unable to find a similar highway to reach voters.
“The Democrats have a last-mile problem when it comes to persuadable voters,” said Dan Pfeiffer, a Democratic strategist who worked on both of President Barack Obama’s campaigns. “Republicans have Fox News and a gazillion Facebook-friendly conservative media, and we as a party have relied anachronistically on the mainstream media.”
Mr. Pfeiffer, who has been briefed on some of the Barometer results, said that the findings, and the expanded use of boosted news, could come at a critical time in the election: With most voters spending more time at home and on screens during the coronavirus outbreak, there is ample opportunity to expand a digital advertising operation.
Pacronym ads have responded accordingly. For most of early April, the group’s ads attacking Mr. Trump’s handling of the coronavirus outbreak highlighted past statements in which he played down the risk of the virus.
But after their testing on the coronavirus was completed in mid-April, the group found that focusing on the plight of health care professionals, rather than on some of Mr. Trump’s earlier comments, had a greater impact.
On April 13, Pacronym switched its large-scale ads on Facebook to news clips of health care professionals running out of personal protective equipment; one clip featuring an NBC report about nurses forced to wear garbage bags to protect themselves reached as many as three million people, according to Facebook’s ad library.
“You’ve got to think about this not as a magic ad,” said David Plouffe, the campaign manager for Mr. Obama in 2008 and a member of the Acronym board. “It’s a constant narrative to people who are really tough to reach, really tough to persuade, and really tough to take action.”