The Internet Research Agency and the Troll-Farm Economy

A building in Saint Petersburg

The Internet Research Agency, sometimes called Glavset internally, ran out of Saint Petersburg from 2013 until it closed in 2023. Court filings and later journalism tied its funding and direction to Yevgeny Prigozhin, the caterer turned mercenary financier who also built the Wagner Group. For its first two years the organization operated in relative obscurity, a fact that says something about how new this kind of operation was: nobody outside Russia had a name for what it was doing, so nobody was looking for it.

That changed in June 2015, when The New York Times Magazine published Adrian Chen's investigation into the agency, the first detailed public reporting on its operations. Chen described a building full of employees writing comments, blog posts, and social media updates on a schedule, targeting Russian domestic audiences and, increasingly, English-language ones. The piece read at the time as a curiosity: a government paying people to argue on the internet. Two years later, testimony before the US Congress would show what that curiosity had grown into.

By 2015, more than 1,000 people reportedly worked inside a single IRA building, running fake accounts across major social networks, discussion boards, and video platforms. That headcount matters as much as anything the agency ever posted. Classical state propaganda, the kind fought over the airwaves and the wire services through the Cold War, ran on a comparatively small number of trained officers working under diplomatic cover or deep alias, because each identity took years to build and each channel of distribution had to be captured or bought outright. The IRA ran on the opposite model: a large, salaried floor of ordinary employees, each producing content in bulk for platforms that would distribute it for free the moment someone hit share.

The economics of a fake American

A Facebook account carrying a stolen or invented American identity costs almost nothing to create and, once it accumulates followers, next to nothing to run. That is the mechanical fact underneath the IRA's return on its spending. Facebook's own account to the US Congress, delivered by General Counsel Colin Stretch to a Senate Judiciary Subcommittee in October 2017, put the scale in numbers a media buyer would recognize: roughly 3,000 IRA-purchased ads reached about 10 million Americans, a paid-media result that would be unremarkable for a well-funded issue campaign.

The number that made headlines was not the paid figure. Stretch testified that once organic sharing and reposting of the agency's pages and posts were counted alongside the paid ads, content traceable to the IRA may have reached approximately 126 million people, over half the American voting-age population. A modest advertising outlay bought the paid distribution; the platform's own architecture, built to reward whatever gets shared, did the rest for free. No propaganda operation before the platform era had that second stage available to it. A pamphlet, a broadcast, or a planted news story had to be re-distributed by other institutions, editors, broadcasters, or a postal system, each of which could refuse. A Facebook post that a stranger finds compelling redistributes itself.

The estimate deserves a caveat the company itself did not fully resolve in public. The path from roughly 29 million people served content directly by IRA-controlled pages to 126 million after shares and reposts depends on how Facebook counted organic reach internally, a methodology it did not publish in full. The narrower, paid-only figure of roughly 10 million is closer to an audited number; the wider figure is the company's own upper estimate of a chain of shares it could not fully trace. Both figures are worth holding, and worth holding differently: one measured, one estimated from measurement.

What ties the ad spend to the headcount is the same economic logic running underneath both. A salaried floor of writers is cheap per unit of output compared with case officers running deniable human assets, and a platform priced for advertisers sells reach by the thousand at a rate any state budget absorbs without strain. Disinformation, on this model, stopped being espionage's expensive cousin and became a line item: staff, ad accounts, and a distribution system somebody else had already built.

What the fake accounts actually said

The content itself was not a single message pushed at a single audience. Facebook told Congress that IRA-linked material touched LGBT issues, race, immigration, and gun rights, running across the ideological spectrum rather than favoring one side of any of those arguments. The company's own characterization was blunt: the material was seemingly intended to amplify societal divisions and pit groups of people against each other.

That description points at something the classical propaganda literature did not have to reckon with. A Cold War broadcast picked a side and argued for it, because the medium, a single transmitter reaching a mass audience at once, only allowed one message per channel. A platform lets an operator run many channels into the same argument at once, some posing as advocates and some posing as opponents, each channel reinforcing the sense that the other side is unreasonable. The goal was not to win an argument on Russia's behalf. It was to make the argument itself less winnable for anyone.

This is the same openness that makes platforms valuable in the first place, held up to different light. The ability of an ordinary person to publish to a mass audience without an editor's permission is what let genuine grassroots movements organize outside legacy media gatekeepers through the 2010s. The IRA's operation shows the other side of that same permission: a well-resourced, patient state actor can occupy the identity of an ordinary person as easily as an ordinary person can, and a platform built to stay neutral about who is speaking has no native way to tell the difference. Liberation and concentration of power turned out to be the same feature, read from two directions.

December 2015 to November 2016

The operation was not a burst of activity timed to Election Day. A January 2017 assessment by the US Intelligence Community, "Assessing Russian Activities and Intentions in Recent US Elections," found that the IRA began advocating for candidate Donald Trump as early as December 2015, months before the first primary ballots were cast in Iowa in February 2016 and nearly a year before the general election that November.

That timing matters for the same reason the headcount matters. A year-long campaign, sustained across primaries, conventions, debates, and the general election itself, is a workforce commitment, not a stunt. It required staff on payroll through an entire election cycle, producing content responsive to each week's news rather than a single scripted burst. The economics that made the ad spend cheap also made the sustained schedule affordable: once the fake identities existed and had followings, keeping them active cost little more than the salaries already being paid.

None of this required the IRA to invent American political division. Immigration, race, guns, and LGBT rights were already the country's live fault lines in 2015. What the operation added was volume and persistence on both sides of each fault line, at a moment when the platforms carrying that volume had not yet built the verification systems, later added under public and regulatory pressure, that might have slowed it down.

The indictment that made it doctrine

On February 16, 2018, a US federal grand jury indicted 13 Russian nationals and three Russian entities, including the Internet Research Agency itself, on charges of interfering with the administration of US elections and political processes. The indictment was not a diplomatic statement or a think-tank report. It was a criminal filing, built on evidence a prosecutor was prepared to defend in court, naming a media operation as a conspirator against a democracy the way a court might name a smuggling ring.

That distinction is the geopolitical hinge of the episode. Before February 2018, a state-run social media operation designed to divide a foreign public sat in an ambiguous category: distasteful, plausibly deniable, and hard to name as an act of conflict rather than an act of speech. A grand jury indictment forced a decision. The US government was now on the record, in a court filing, treating platform-based influence operations as an instrument used against it, alongside espionage and cyberattack, rather than as background political noise that happened to have a Russian accent.

The effect reached beyond the United States. Governments and alliances that had spent 2016 and 2017 debating whether disinformation was a real security problem or a media panic now had a working legal precedent for the harder version of the question: not whether foreign influence content existed, but whether it counted as the kind of interstate act that intelligence services, election commissions, and defense alliances exist to counter. The IRA case became the reference point cited whenever a government since has had to explain, in public, why it treats a feed manipulation campaign the way it once treated a spy ring.

Shutdown, and what outlived it

The Internet Research Agency shut down on July 1, 2023, in the direct aftermath of the Wagner Group's June 2023 rebellion against Russia's military leadership. Wagner and the IRA shared a patron in Prigozhin; when his position collapsed, so did the organization he had funded. The agency's end was a political event inside Russia, not evidence that the tactic it had industrialized had stopped working, or stopped being used elsewhere.

What the IRA proved does not require a building in Saint Petersburg to repeat. A workforce that costs little, operating fake identities on platforms that redistribute engaging content for free, at a scale that a February 2018 indictment made a government treat as a matter of national defense: that combination is available to any state or organization willing to staff it, and the platforms that made it possible in 2015 have only multiplied since.

The unresolved problem underneath the story is also one this publication tracks in its own present-day form. The IRA exploited a gap that had nothing to do with the content of any single post: a feed had no reliable way to tell an authentic voice from a manufactured one, so it treated both as equally worth amplifying. A system assembling an answer today, drawing on whatever sources it can read and has reason to trust, faces a version of the same gap. Credibility, one of the four dimensions this publication measures as machine readiness, is the current name for the problem the IRA spent a decade proving was worth exploiting.