It’s not every day that a former work colleague gets retweeted by the president of the United States.
Last Friday, Rob Goldman, a vice president inside Facebook’s Ads team, rather ill-advisedly published a series of tweets that seemed to confirm the Trump administration’s allegations regarding the recent indictments of 13 Russian nationals by Special Counsel Robert Mueller. To wit, the tweets said that the online advertising campaign led by the shadowy Internet Research Agency was meant to divide the American people, not influence the 2016 election.
You’re probably skeptical of Rob’s claim, and I don’t blame you. The world looks very different to people outside the belly of Facebook’s monetization beast. But when you’re on the inside, like Rob is and like I was, and you have access to the revenue dashboards detailing every ring of the cash register, your worldview tends to follow what advertising data can and cannot tell you.
From this worldview, it's still not clear how much influence the IRA had with its Facebook ads (which, as others have pointed out, is just one small part of the huge propaganda campaign that Mueller is currently investigating). But no matter how you look at them, Russia’s Facebook ads were almost certainly less consequential than the Trump campaign’s mastery of two critical parts of the Facebook advertising infrastructure: The ads auction, and a benign-sounding but actually Orwellian product called Custom Audiences (and its diabolical little brother, Lookalike Audiences). Both of which sound incredibly dull, until you realize that the fate of our 242-year-old experiment in democracy once depended on them, and surely will again.
Like many things at Facebook, the ads auction is a version of something Google built first. As on Google, Facebook has a piece of ad real estate that it’s auctioning off, and potential advertisers submit a piece of ad creative, a targeting spec for their ideal user, and a bid for what they’re willing to pay to obtain a desired response (such as a click, a like, or a comment). Rather than simply reward that ad position to the highest bidder, though, Facebook uses a complex model that considers both the dollar value of each bid as well as how good a piece of clickbait (or view-bait, or comment-bait) the corresponding ad is. If Facebook’s model thinks your ad is 10 times more likely to engage a user than another company’s ad, then your effective bid at auction is considered 10 times higher than a company willing to pay the same dollar amount.
A canny marketer with really engaging (or outraging) content can goose their effective purchasing power at the ads auction, piggybacking on Facebook’s estimation of their clickbaitiness to win many more auctions (for the same or less money) than an unengaging competitor. That’s why, if you’ve noticed a News Feed ad that’s pulling out all the stops (via provocative stock photography or other gimcrackery) to get you to click on it, it’s partly because the advertiser is aiming to pump up their engagement levels and increase their exposure, all without paying any more money.
During the run-up to the election, the Trump and Clinton campaigns bid ruthlessly for the same online real estate in front of the same swing-state voters. But because Trump used provocative content to stoke social media buzz, and he was better able to drive likes, comments, and shares than Clinton, his bids received a boost from Facebook’s click model, effectively winning him more media for less money. In essence, Clinton was paying Manhattan prices for the square footage on your smartphone’s screen, while Trump was paying Detroit prices. Facebook users in swing states who felt Trump had taken over their news feeds may not have been hallucinating.
(Speaking of Manhattan vs. Detroit prices, there are some (very nonmetaphorical) differences in media costs across the country that also impacted Trump’s ability to reach voters. Broadly, advertising costs in rural, out-of-the-way areas are considerably less than in hotly contested, dense urban areas. As each campaign tried to mobilize its base, largely rural Trump voters were probably cheaper to reach than Clinton’s urban voters. Consider Germantown, Pa. (a Philly suburb Clinton won by a landslide) vs. Belmont County, Ohio (a rural county Trump comfortably won). Actual media costs are closely guarded secrets, but Facebook’s own advertiser tools can give us some ballpark estimates. For zip code 43950 (covering the county seat of St. Clairsville, Ohio), Facebook estimates an advertiser can show an ad to about 83 people per dollar. For zip code 19144 in the Philly suburbs, that number sinks to 50 people an ad for every dollar of ad spend. Averaged over lots of time and space, the impacts on media budgets can be sizable. Anyway …)
The above auction analysis is even more true for News Feed, which is only based on engagement, with every user mired in a self-reinforcing loop of engagement, followed by optimized content, followed by more revealing engagement, then more content, ad infinitum. The candidate who can trigger that feedback loop ultimately wins. The Like button is our new ballot box, and democracy has been transformed into an algorithmic popularity contest.
But how to trigger the loop? For that, we need the machinery of targeting. (Full disclosure: I was the original product manager for Custom Audiences, and along with a team of other product managers and engineers, I launched the first versions of Facebook precision targeting in the summer of 2012, in those heady and desperate days of the IPO and sudden investor expectation.)
Despite folklore about “selling your data,” most Facebook advertisers couldn’t care less about your Likes, your drunk college photos, or your gossipy chats with a boyfriend. What advertisers want to do is find the person who left a product unpurchased in an online shopping cart, just used a loyalty card to buy diapers at Safeway, or registered as a Republican voter in Stark County, Ohio (a swing county in a swing state).
Custom Audiences lets them do that. It’s the tunnel beneath the data wall that allows the outside world into Facebook’s well-protected garden, and it’s like that by design.
Browsed for shoes and then saw them on Facebook? You’re in a Custom Audience.
Registered for an email newsletter or used your email as login somewhere? You’re in a Custom Audience.
Ordered something to a postal address known to merchants and marketers? You’re definitely in a Custom Audience.
Here’s how it works in practice:
A campaign manager takes a list of emails or other personal data for people they think will be susceptible to a certain type of messaging (e.g. people in Florida who donated money to Trump For America). They upload that spreadsheet to Facebook via the Ads Manager tool, and Facebook scours its user data, looks for users who match the uploaded spreadsheet, and turns the matches into an “Audience,” which is really just a set of Facebook users.
Facebook can also populate an audience by reading a user’s cookies—those digital fragments gathered through a user’s wanderings around the web. Half the bizarre conspiracy theories around Facebook targeting boil down to you leaving a data trail somewhere inside our consumer economy that was then uploaded via Custom Audiences. In the language of database people, there’s now a “join” between the Facebook user ID (that’s you) and this outside third-party who knows what you bought, browsed, or who you voted for (probably). That join is permanent, irrevocable, and will follow you to every screen where you’ve used Facebook.
The above is pretty rudimentary data plumbing. But only when you’ve built a Custom Audience can you build Lookalike Audiences— the most unknown, poorly understood, and yet powerful weapon in the Facebook ads arsenal.
With a mere mouse click from our hypothetical campaign manager, Facebook now searches the friends of everyone in the Custom Audience, trying to find everyone who (wait for it) “looks like” you. Using a witches’ brew of mutual engagement—probably including some mix of shared page Likes, interacting with similar News Feed or Ads content, a score used to measure your social proximity to friends—the Custom Audience is expanded to a bigger set of like-minded people. Lookalikes.
(Another way to picture it: Your social network resembles a nutrient-rich petri dish, just sitting out in the open. Custom Audiences helps mercenary marketers find that dish, and lets them plant the bacterium of a Facebook post inside it. From there, your own interaction with the meme, which is echoed in News Feed, spreads it to your immediate vicinity. Lookalike Audiences finishes the job by pushing it to the edges of your social petri dish, to everyone whose tastes and behaviors resemble yours. The net result is a network overrun by an infectious meme, dutifully placed there by an advertiser, and spread by the ads and News Feed machinery.)
We’ve all contributed to this political balkanization by self-sorting (or being sorted by Facebook) into online tribes that get morphed into filter bubbles, which are then studiously colonized by commercial memes planted and spread there by a combination of Custom and Lookalike Audiences. One of the ways the Trump campaign leveraged Lookalike Audiences was through its voter suppression campaigns among likely Clinton voters. They seeded the Audiences assembly line with content about Clinton that was engaging but dispiriting. This is one of the ways that Trump won the election, by the very tools that were originally built to help companies like Bed Bath & Beyond sell you towels.
Unsurprisingly, the Russians also apparently made use of Custom Audiences in their ads campaign. The unwary clicker on a Russian ad who then visited their propaganda site suddenly could find yet more planted content in their Feed, which could generate downstream engagement in Feed, and thus the great Facebook wheel turned. The scale of their spend was puny, however, a measly $100,000, which pales in comparison to the millions Trump spent on online advertising.
The above isn’t mere informed speculation, the Trump campaign admitted to its wide use of both Custom and Lookalike audiences. There seems to be little public coverage of whether the Clinton campaign used Facebook Ads extensively, but there’s no reason to think her campaign did not exploit the same tools.
“I always wonder why people in politics act like this stuff is so mystical,” Brad Parscale, the leader of the Trump data effort, told reporters in late 2016. “It’s the same shit we use in commercial, just has fancier names.”
He’s absolutely right. None of this is even novel: It’s merely best practice for any smart Facebook advertiser. Custom Audiences was launched almost six (!) years ago, marketed publicly at the time, and only now is becoming a mainstream talking point. The ads auction has been studied by marketers and academics for even longer. The only surprise is how surprising it can still seem to many.
If we’re going to reorient our society around Internet echo chambers, with Facebook and Twitter serving as our new Athenian agora, then we as citizens should understand how that forum gets paid for. Rarely will the owners of that now-privatized space deign to explain how they’re keeping the lights on. Plotting Russians make for a good story, and external enemies frequently serve an internal purpose, but the trail of blame often leads much closer to home. It’s right there, topped by a big, blue bar on our smartphone screens, and could very well be how you arrived at what you’re reading right now.
Update (February 27, 2018): In an unusual move, Andrew 'Boz' Bosworth, former VP of Facebook Ads, posted average CPMs for both the Clinton and Trump campaigns this afternoon. The figures are national averages over time, and while they fluctuate wildly, they mostly show the Trump campaign paying more on a CPM basis than Clinton. While interesting, and the transparency of Facebook is admirable, the data only refute the rather strong statement that Trump always and everywhere paid less. By and large, these data do not confirm or deny the hypotheses contained in this piece.
The data that Facebook needs to show us are average CPMs broken down by targeting type, action type (e.g., clicks or likes), and geography. The first two would help distinguish direct-response campaigns, which typically are precision targeted and high CPM, from more brand-style ad campaigns that are broadly targeted and low CPM. Combining the data from both styles of campaign---which broadly define the two types that advertisers undertake---can be very deceptive, and the two campaign types need to be judged separately.
Furthermore, a breakdown by geography would help determine whether another assertion made in this piece is correct: That Trump paid less to mobilize his base than Hillary. Obviously, combining data nationwide makes this very hard to figure out.
Reportedly, Facebook has asked the campaigns to be more forthcoming with data. As it's in both those campaigns' interests at this point, one can only hope they do so. As we used to say at Facebook: "Data wins arguments."
- Rob Goldman, VP of ads at Facebook, published a tweetstorm on Friday appearing to confirm the Trump administration’s allegations around the ongoing Muller investigation ... and he did so without clearing his contributions with his employer.
- No, Facebook isn't eavesdropping on you through your phone to better target you with ads. It doesn't have to.
- To fix its toxic ad problem, Facebook will have to undergo a massive cultural shift.
Photograph by WIRED/Getty Images
