Donald Trump’s Facebook campaign was crucial to his 2016 success, but two years later the nuts and bolts of the operation that helped sweep Trump into the White House remain hard to come by.

Now new data has opened a small window into Trump’s social-media machinery, and in particular the role played by the now-defunct Cambridge Analytica, a U.K. digital black-ops firm that collapsed this year following revelations that it acquired and used detailed Facebook profile information on 87 million people without their knowledge.

In the wake of the privacy scandal, Trump and Republican National Committee officials have sought to distance the campaign from Cambridge Analytica. But questions persist. Last year, Robert Mueller sought emails between Cambridge Analytica and Trump campaign staffers, and in March Mueller’s team questioned former campaign officials about the U.K. firm’s work for Trump.

A similar probing is playing out in the Senate Intelligence Committee's Russia investigation. As The Daily Beast reported Thursday, Senate investigators interrogated Steve Bannon about Cambridge Analytica behind closed doors last month.

Public statements and insider accounts have painted a muddled and contradictory picture on the key question of whether Trump’s Facebook campaign targeted voters using Cambridge’s vast store of dubiously acquired data, once described by the company as containing 4,000 data points on some 230 million Americans.

Now a New York digital-marketing consultant has unearthed a trove of digital artifacts from Trump’s social-media campaign that provides the first hard evidence that Team Trump made continuous use of audience lists created by Cambridge Analytica to target a portion of its “dark ads” on Facebook. The ads were deployed from July 2016 through the end of the election—and beyond, to the inauguration in January 2017.

A Trump 2016 campaign official confirmed those findings to The Daily Beast, but claimed that Cambridge Analytica built the audience lists from the RNC’s database of voters and not its internal data store.  

At issue are the “dark ads” or “dark posts” that underpinned Trump’s 2016 social-media campaign. Until recently, advertisers could use Facebook’s precision-targeting tools to run ads that nobody except the targets would ever see, evading broader scrutiny. Those hidden ads have been the center of attention since a 2016 Bloomberg News story reported that Trump was using dark ads to stealthily target black Americans and other likely Clinton voters and urge them to stay home on Election Day.

The Trump campaign has disputed the story, and no such ads have ever surfaced publicly. But few dispute that Facebook’s system carried the potential for such abuse.

Under pressure, this year Facebook created a tool allowing any user to look up past and current political ads regardless of the targeting, but the tool does not extend back to the 2016 election, and Facebook has resisted calls to make ads from the presidential race public.

But some traces remain.

Emily Las, a digital marketer and former VP at MasterCard, has spent the last year extracting remnants of Trump’s Facebook campaign through and beyond Election Day 2016. So far, her work has unearthed more than 1,200 tracking links for different Trump ads, and live content for hundreds of the ads.

That’s a minuscule slice of a campaign that reportedly launched 5.9 million ads, but every tracking link carries a DNA strand of data about the overall campaign in the form of UTM (“Urchin Tracking Module”) codes.  These codes, long cryptic sequences of characters, allow an advertiser to track an ad’s performance.

The UTM parameters are a kind of digital hobo’s code, obscure to the consumers who click on them, but mostly legible to anyone steeped in digital marketing, as Las has been for years. She’s been dissecting her collection of 1,200 links under a microscope and breaking out every data point in a meticulous spreadsheet. “When the RNC took over, you can tell it becomes a more sophisticated operation,” she said. “They’re tracking every little element of the ad."

Over time, Las has developed a deep picture of some aspects of Trump’s Facebook campaign, details like which technology partners Trump used at various points in the campaign, and what the intended purpose of each ad was.

The ads in her collection were predominantly coded for fundraising, under a handful of industry standard subcategories. “Retention” ads, for example, were deployed to encourage past donors to donate again. These ads frequently offered enticements like a sweepstakes drawing for lunch or dinner with Trump, a place on a “donor wall” in Trump Tower, or simply a promise that every donor’s name would be inscribed on a list of patriotic supporters that would be delivered into Trump’s hands.

Another subcategory called “prospecting” refers to ad buys designed to lure targets into providing their email address for the campaign’s list, which is now a valuable commodity in itself. Las’ data suggests ersatz petition drives were Trump’s primary hook for this data collection, and the campaign’s prospecting continued well after Election Day. In the last year, prospecting ads have sent users to petitions to end “chain migration” and support Judge Brett Kavanaugh’s nomination to the Supreme Court.

The most intriguing tidbits in the UTM codes are those that categorize the Facebook “audience” being targeted—meaning a set of users picked to receive a certain ad, or to serve as a template for Facebook’s “lookalike audience” matching feature. In about 10 percent of her tracking links, the audience is described as “CambridgeAudience” or “CambridgeAnalytica.”

Cambridge Analytica was a data-mining and influence firm partially bankrolled by the Mercer family and co-founded by former Trump consigliere Steve Bannon, who also served on the board. The company touted its ability to influence behavior using an electronic store of information on consumers, including 4,000 data points on 230 million Americans.

The Trump campaign paid Cambridge Analytica nearly $6 million in 2016. Campaign officials have said $5 million of that money was earmarked for TV ads, and that the company itself was only paid about $800,000.

Whatever the amount, Cambridge Analytica fell into disrepute this year over new revelations about its sneaky acquisition of profile data on 87 million Facebook users, which it was allegedly using to build psychological models. At the same time, a hidden-camera sting by Britain’s Channel 4 caught company executives boasting of the firm’s skill at executing covert, untraceable election-interference ops using a combination of data analytics, fake news, false-front organizations, and timeworn dirty tricks like hiring prostitutes to lure a rival candidate into a honey trap.

The Trump campaign was drawn into the controversy by Cambridge CEO Alexander Nix, who was heard on one of the undercover videos boasting about his company’s work for Trump and effectively claiming credit for the Apprentice star’s election victory.

Cambridge Analytica, Nix said, “did all the research, all the data, all the analytics, all the targeting, we ran all the digital campaign, the television campaign, and our data informed all the strategy.”  

Parscale, who’d hired the firm, called out Nix’s boast as “false and ridiculous,” and in the months since he and other Trump and RNC officials have been busy downplaying Cambridge Analytica’s work for the campaign.

In Parscale’s account, he was never interested in Cambridge Analytica’s vaunted databases, just its people. In particular, Parscale had been eager to poach the firm’s product chief, data guru Matt Oczkowski. When he couldn’t hire Oczkowki away from the firm, he resorted to engaging the company as a kind of high-end temp agency just to get Oczkowski and his team to San Antonio.

“I asked them for an employment contract and I hired them for staff only,” Parscale said in an interview with PBS Frontline last month. “So each one of the payments between then and Election Day was for staff only.”

As for what Cambridge’s team actually did, Parscale said they “mainly ran polling, visualization, and support staff for all of the things we needed to do.” He emphasized again that he didn’t “hire Cambridge Analytica… for any of their data.”

This portrait of the Cambridge team as de facto Trump employees doesn’t fit easily with the evidence uncovered by Las, which shows a steady stream of ads targeted to “Cambridge” audiences. The Cambridge references are found in 120 out of 1,200 ad links, spread evenly from July 2016 to January 2017, when a slew of retention ads used inauguration tickets and memorabilia as a dangle.

An official from Trump’s 2016 campaign, speaking to The Daily Beast on condition of anonymity, confirmed that these links were for ads targeted by Cambridge Analytica. But the official said the audience lists were built in San Antonio by Cambridge workers who didn’t use their company’s data. Instead, the underlying “first person” data came from the RNC’s list and the Trump campaign’s internal database, code-named Alamo.

Cambridge’s Oczkowski, who’s now working for Trump 2020 through his own firm, has backed up that claim following Cambridge’s scandals. Prior to those scandals, though, he explicitly described Cambridge’s dataset as part of the Trump operation.

In a December 2016 Google roundtable discussion, Oczkowski expounded on the complexity of bootstrapping Trump’s social-media campaign from three disparate data sources, including Cambridge Analytics.

“On the targeting piece, we’re talking about building a database, working with the RNC, the Alamo database, and then leveraging Cambridge’s database,” he said. “Combining those three things together, building partisanship models, 12 issue sets, the basic building blocks you need from a campaign.”

If Oczkowski misspoke, the other Trump campaign officials on the panel didn’t say anything, including the RNC’s Gary Coby, who served as Trump’s digital ads and fundraising director, and Parscale himself.

The shifting storylines surrounding the most successful social-media ad campaign in history is what sent Emily Las digging for clues last year. “I’m still fixated on understanding 2016,” she told The Daily Beast. “I really wanted to know what they did, and the more information that came out, the more that it didn’t make sense to me. And I thought, nobody’s going to tell me, so I’m just going to build it up.” Las was particularly intrigued by the 2016 Bloomberg News profile of Brad Parscale. Bloomberg reported that Parscale was operating three highly targeted voter-suppression campaigns discouraging white liberals, young women, and black Americans from turning out for Hillary Clinton on Election Day. The effort reportedly included a dark post on Facebook consisting of a South Park-style animation titled “Hillary Thinks African Americans are Super Predators.”

In addition to the tracking links, Las has discovered and documented hundreds of Trump 2016 dark posts that are still alive deep in Facebook’s servers. Hidden outside the reach of Facebook’s search engine and not displayed on any timeline, the ads, including nearly 500 distinct videos, are findable only by those who know the link.

So far there are no surprises in the ads—they don’t differ in tone from the Trump campaign’s overt messaging or Donald Trump’s public statements at rallies. Some, however, illustrate the A/B testing that Trump’s digital team touted in interviews.

For example, in mid-August 2016, the campaign promoted a seven-day “Trump train” fundraising drive. Two of the unearthed versions of the ads feature the text, “$7 million. In 7 days. Welcome to the #TrumpTrain. Fuelled by... America.” But one version of the ad is accompanied by jazzy, percussive music, a voice-over, and stock footage depicting the “Trump train” as a sleek, fast-moving passenger train. A second version has brighter music, no voice-over, and the Trump train is now a freight train chugging unhurriedly through the heartland.

Similarly, a post-election ad from Oct. 19 of last year urged Trump supporters to enter a raffle and win an all-expenses paid dinner with the president. But while some supporters were presented with a vanilla ad, others saw the same video with a “BREAKING NEWS” banner and a fake countdown timer, warning that less than five minutes remained to enter the drawing, which was actually set to close the next day.

Las, who is not a Trump supporter, forced herself to watch every variation of each of the recovered ads. “It’s like watching one long boring infomercial. This is just horrible quality. It’s cheesy."

But in all her sleuthing, Las hasn’t turned up any evidence of the supposed dark-ad voter-suppression campaign that first sent her down the tracking-link rabbit hole.

The official from Trump’s 2016 campaign said there’s a reason for that: Trump had no voter-suppression ads. Bloomberg’s reporting, the official said, was based on a proposal floated by an in-house vendor but never approved nor implemented.

On this point, Las finds herself in rare agreement with the Trump campaign. “There’s nothing in here that feels like it’s about suppression,” she said. After a year of searching, “I don’t think it exists.”