TikTok creators are telling you life is unaffordable. They’re being paid to.

The neutral-looking accounts were built first

Account creation is not evenly spread. It arrives in three bursts — and the accounts presenting as local news feeds predate the personas that do the talking.

Wave 0 builds an audience on affordability and local-news framing. Wave 1 drops the talking-head personas into that ready-made follow graph — this is the wave the video came from. Wave 2 is still running, and was still running as this report was written: @chels.chats was created three days before collection with zero posts and already follows eight network members, and @rantswithbec — "Georgia girl with a lot to say · professional yapper, casual ranter" — has zero followers, zero posts and fourteen follows, five of them confirmed members. An empty persona, wired into the graph before it has said anything.

The chart has a blind spot, and it is worth naming. kelsie.explains was registered in May 2021 — but every one of its eight posts landed between 11 and 15 August 2026. It is a dormant account switched on for the campaign, and creation-date analysis cannot see it. Anyone reactivating old accounts rather than making new ones is invisible to this exhibit — and kelsie.explains is no longer the only one. ganewswithparis, a Georgia account with eleven inbound follows from confirmed members, was registered in January 2024. Several of the one hundred and thirteen accounts here pre-date the waves entirely and would never have surfaced through creation-date analysis — two dormant reactivations, plus mads_ventures, an established 2019 creator with 410 videos who simply took the brief on the account they already had. One hundred and thirteen is therefore a floor, not a count.

How the waves were identified

A fourth wave, found after the first version of this report. Enumerating the Texas and Michigan cells surfaced a provisioning cluster the original three-wave model did not contain: fourteen accounts created between 28 April and 22 May, four of them on a single pair of days. Concentrated in Texas but reaching Michigan, Iowa and California. The wave is plotted above and the dated list is in Annex E.

Why this wave was missed the first time

Why it matters that this was missed the first time. The wave model in this section was built on the accounts known at the time. It described the network as provisioned in three bursts. A single further round of enumeration added a fourth burst that is larger than two of the original three. The waves are real — the clustering is far too tight to be coincidence — but the number of waves is a function of how much of the network has been found, and should be read that way.

One formula: a first name, and a word meaning "to talk"

Forty of the 113 handles are built from the same template — a name plus an informal verb for speech. The verbs are interchangeable, and one operator proves it: nimaiyaps and nimaitalks are the same person, running two accounts whose only difference is which synonym they picked.

Handle shape alone proves nothing — plenty of real people use these constructions, and a search for "rants with" returns mostly fandom accounts. It carries weight here only because every account below also sits inside the same follow graph, the same creation window and the same script.

Scroll table sideways on narrow screens

Two products, sold as one grassroots

Post captions in this network are bare hashtag strings and many videos contain no speech at all — the message lives in text burned into the frame. To read it at scale, the text was read off the video frame itself. 9,261 posts across 105 accounts. The result is the most important finding in this report.

Michigan is the finding this table was missing. Half of Michigan's described posts — 50% — mention voting, elections, ballots or the midterms. That is the highest election rate anywhere in the network, above New York's 29%, and it sits on a cell of only three accounts. Its partisan-attribution rate, 39%, is second only to New York. Michigan is a presidential swing state, and it is being run as the most electorally explicit cell in the operation.

Texas arrived as a tax campaign. The newest and second-largest cell leads the network on taxes (32%) and runs 21% partisan attribution — unlike Georgia, North Carolina, Pennsylvania and Wisconsin, which are almost purely economic. Iowa, on a single account and only 78 posts, is 56% wages. Both read as narrower, more specific products than the general-grievance cells.

How to read the small cells

Read the small cells carefully. Maine rests on 24 campaign posts and Iowa on 78, against 1,711 for Georgia and 1,240 for Texas. Percentages on that base move several points if a handful of posts are reclassified, which is why Maine and Iowa are excluded from the chart above. Michigan, at 364 posts, is small but not marginal: the direction of its finding is robust, the precise figure less so.

Why these rates changed, and what the old ones measured

Why these rates changed from an earlier version. An earlier table scored each cell only on posts carrying a machine-readable scene description. That instrument is unusable for comparison: its coverage runs from zero percent before 2026 to 86% by August, so it ranked cells by how thoroughly they happened to be collected. Every rate here is now computed on caption plus on-screen text read from the video frame, which covers 99.5% of the catalog and treats a September 2025 post exactly like an August 2026 one. New York fell from 75% to 61%, Michigan from 39% to 30%, Texas from 25% to 21% and Ohio from 24% to 21% — the magnitudes moved, the ordering did not.

The partisan row is the story, and it is not a clean split. Five cells carry real political attribution — New York 61%, Michigan 30%, Iowa 24%, Ohio 21%, Texas 21%. Four carry next to none — Georgia 8%, and North Carolina, Pennsylvania and Wisconsin at one percent or below. The same operator is running an attack product in some states and pure economic anxiety in others, in the same costume, at the same time. An earlier version of this report described that as two products. With ten cells measured it looks more like a dial: the partisan content is turned up or down state by state, and the states where it is turned up are not the states where the network is largest.

The localization is real, and it is granular. Data centers appear in 17% of Ohio posts and 14% of Georgia's — against two posts out of 781 in Pennsylvania, and none at all in Wisconsin or New York. Each state's own utility gets named: FirstEnergy and AEP in Ohio, Duke Energy in North Carolina, Alliant in Wisconsin. This is precisely the capability the operator advertises — "a state, city, congressional district" — executed at the level of which power company sends your bill.

The strand that is not about groceries

One topic in this network is not a cost-of-living theme at all, and it behaves differently from everything else: Flock, the automated license-plate-reader camera system that towns buy for their police departments. 39 posts across 20 accounts — Ohio, Georgia, North Carolina and the ambassador cell — between 17 June and 15 August 2026, carrying 4.5 million views. Ohio runs it hardest, at 19 posts.

It matters for three reasons. First, the framing: almost every one of these posts routes surveillance back through the household budget — "but we have millions for flock camera's", "Georgia homeowners are now being asked to pitch in to pay for", "how am i supposed to afford my college tuition AND rent". The grievance is not privacy. It is money spent on this instead of on you.

Second, the repetition. newswithchloe posted the identical caption — "we need to fix our priorities #ohiotok #ohiopolicy #flock" — five times between 11 and 14 August. Organic opinion does not duplicate itself word for word five times in four days; scheduled content does.

Third, the link to the paid program is explicit. kamyapss, one of the eight recruiter accounts, posts Flock content that carries the recruitment pitch in the same caption — "what do you think about this? and if you want to be paid to…". A topic being pushed by an account whose function is to sign up paid creators is not a topic that arrived organically.

Why this is flagged rather than resolved. Flock is a live local-government controversy in all three states, and ordinary people post about it constantly. The finding here is not that concern about license-plate readers is manufactured — it plainly is not. It is narrower and more specific: a coordinated paid network picked this topic up across four cells inside eight weeks, framed it as a budget grievance rather than a civil-liberties one, and ran it through the same undisclosed personas as the rent and grocery content. Whether that reflects a client brief or creators following a trend is not established from outside.

The caption is deliberately empty

There is a habit running through this network that is easy to miss and hard to explain innocently. The caption very often says nothing at all, and the entire argument is burned into the video as on-screen text.

The clearest case is also the network’s biggest post. A video on mads_ventures that reached 4.2 million views carries the caption “yall we HAVE to stand up” — seven words that name no subject. Read the frame instead and the text across the top of the video says “flock is evil”. The post is about automated license-plate surveillance. Nothing in the caption tells you that.

It is not a one-off. In May 2026 talkingwithh published a run of videos captioned only “Crazy” followed by a block of place-and-industry hashtags — #texas #dallas #gas #gasstation #oil — repeated across separate posts on the 13th, 14th, 15th, 18th, 19th, 25th and 28th. The caption is inert every time. The on-screen text is where the claim lives: that people move to Texas for the low taxes and find the costs waiting for them somewhere else.

One caption, three arguments. Every one of these posts is captioned “Crazy #texas #dallas #gas #gasstation #oil” — identical, and empty. The hashtags do the geographic targeting; the claim is burned into the frame where no caption-based system reads it.

The effect is that the message and the label point in different directions. The hashtags do the targeting — a city, a state, an industry — while the caption stays content-free, and the substance is placed where only a viewer sees it. Automated systems that read captions do not see the claim. Neither does a researcher who audits captions, which is the ordinary way large samples of social media get read.

Measured across the whole catalog, the caption hides a sixth of this operation. Classifying every post on its caption alone finds 6,499 campaign posts and 52.3 million views. Classifying the same posts on caption and on-screen text finds 7,519 posts and 63.4 million views. The difference — 1,020 posts carrying 11.1 million views, 17.5 percent of the campaign — is content that is unmistakably on-brief once you read the frame and invisible if you read only what was typed into the caption box.

That gap is not evenly spread. It is concentrated on the highest-performing posts, because the videos built to travel are the ones given a bare caption and a block of location hashtags. The single largest post in the network sits inside it.

This report initially made exactly that mistake, and says so. An earlier pass classified posts by caption and hashtags. On that reading, the 4.2-million-view surveillance video was not affordability content and the talkingwithh gas series was not cost-of-living content, because their captions do not contain the words. Reading the frame reversed both. Every reach and theme figure in this report is now built on the on-screen text as well as the caption — the method is set out in Annex A — and the direction of the earlier error is worth stating plainly: caption-only reading always undercounts this network, never overcounts it.

What is missing from all 9,261 posts

There is almost no ask. Across 9,261 posts there is, with the two exceptions set out in §06, no candidate to vote for, no bill to support, no petition to sign, no organization to join, no march to attend. Not one of the 113 accounts carries an outbound link in its bio. No donate page, no campaign site, no shop. Four publish a contact email address and nothing else — one of which is how two of the personas were tied to the same operator.

That absence is doing work. Content that assigns blame without requesting an action is not organizing and does not have to behave like it: it never has to name a client, file a disclosure, or account for a dollar. It leaves an impression rather than a demand — that life is unaffordable, that it is someone's fault, and that nothing can be done. Whatever else this is, it is not a campaign asking people to act. It is a campaign asking them to feel a certain way, at scale, without telling them who is paying.

How theme rates are calculated

Method. Theme rates are the share of a cell's campaign posts whose caption or on-screen text matches a keyword set for that theme; posts can match more than one. On-screen text is read by optical character recognition of the video's cover frame, available for 99.5% of the catalog — see Annex A. "Partisan attribution" counts any mention of a party, a named official, MAGA, or voting and elections. Maine (24 posts) and Iowa (78) are excluded from the chart as too thin to compare against cells of a thousand.

This is not social media. It is a delivery system wearing its clothes.

The central point of this report is easy to lose in the account counts, so state it plainly.

Ordinary influencer marketing pays an existing person with an existing audience and an existing relationship with that audience. The follower chose to follow them. Whatever else is wrong with it, there is a real human relationship being monetized, and the audience can see the account's history.

That is not what happened here. The accounts were built for the campaign. Twenty-four personas created in three waves across seven weeks, each spun up on two platforms simultaneously, each with no history, no prior audience and no relationship to monetise. The persona is not a person who was hired — it is infrastructure that was manufactured, and the human in front of the camera is the last component fitted.

The people delivering the lines may sincerely believe every word. That is not the issue, and this report makes no claim about their convictions. The issue is that they were paid to say it, handed "content inspo and viral hooks" to say it with, and placed on accounts constructed so the audience would read them as a neighbour rather than a channel. Belief is irrelevant to the deception; the concealment of the arrangement is the deception.

The effect at scale is not persuasion so much as saturation. The founder's own description — "volume is the key mechanism," "a lot of authentic shots on goal," 70+ posts in a week, 2–3 per creator per day — is a strategy of flooding a local feed with scripted, prepackaged goods until the synthetic version of a conversation crowds out the real one. Someone in Columbus scrolling their local tag does not encounter one paid message; they encounter a manufactured consensus.

And the counterfeit is deliberately hard to tell from the real thing, because the counterfeit is sold on being hard to tell from the real thing. That is what "beat paid advertising on both trust and cost" means. The trust premium is harvested from the audience's belief that nobody is being paid.

9,261 posts. Almost every disclosure came from a recruiter.

The full corpus — 9,261 posts across 105 accounts — was searched for any disclosure signal at all: a Drumbeat mention, #ad, #sponsored, #paidcreator, or a paid-partnership tag. The distribution is perfect, and that is what makes it damning.

Almost no persona discloses: 55 posts out of 8,463, and 42 of those sit on one account. Every other disclosure in this network — 126 posts in all — comes from the handful of accounts whose job is recruiting more creators. On those, the tag is near-universal: nimaiyaps carries it on sixteen posts out of sixteen. The accounts doing the persuading carry it on none.

The disclosure evidence in full

A blanket absence could be explained by ignorance — nobody told the creators the rule existed. That explanation is not available. This operation has a disclosure convention, applies it consistently when talking to creators, and applies it zero times when talking to voters.

This is the finding, and it is not the one I expected. A blanket absence of disclosure could be explained by ignorance — nobody told the creators the rule existed. That explanation is not available here. The operation has a disclosure convention, #DrumbeatAmbassador, and applies it consistently on the two recruiter accounts: sixteen posts out of sixteen, eleven out of twelve. On the ninety-three accounts doing the actual persuasion, the ambassador tag appears zero times in 8,463 persona posts — while a paid-political disclosure appears on nineteen, all from one account. The people who built this know exactly how to disclose. They disclose on the accounts that recruit, and not on the accounts that talk to voters.

Note also that neither disclosing account ever flips TikTok's branded-content toggle — the mechanism the platform actually built for this. A hashtag is not the toggle, and a hashtag on the recruiter is not disclosure on the persona.

TikTok's rules

TikTok requires creators to switch on the branded content toggle whenever a post is made under any commercial relationship. Separately and more importantly, TikTok prohibits political advertising and political branded content outright — the toggle cannot lawfully be used for political or issue content on that platform at all. A campaign of paid, briefed posts attacking a political party therefore fails the policy twice over: undisclosed if the toggle is off, and prohibited if it were on. On its face, this network breaches TikTok's Branded Content Policy. That conclusion needs no regulator and no interpretation of statute — it follows from the platform's published rules and the posts above.

Instagram's rules — breached in exactly the same way

This is not a TikTok story with an Instagram footnote. Fifty-six of the 113 handles are verified cross-platform mirrors, most under the identical name, and twenty of those publish a post count within one of their TikTok count. The same briefed content, from the same purpose-built personas, is being pushed to Meta's audience at the same volume — jazzmyneohio with 54 Instagram posts, newswithjob with 53, oliviainohio with 70. The operator's own instruction sheet says it outright: "Set up new TikTok and Instagram accounts." Instagram is not incidental to this campaign; it is half of it.

Meta's Branded Content Policies require that any content posted in exchange for value carry the "Paid partnership with" label, applied through Meta's branded-content tool — it is mandatory, not optional, and the obligation sits on the creator. Content concerning social issues, elections or politics carries a further regime: advertiser authorization, identity verification and a "Paid for by" disclaimer. Meta's Inauthentic Behavior policy separately prohibits misleading people about the origin or purpose of content.

The Instagram audit sharpens this. The two ambassador accounts disclose there too — kamyapss' Instagram bio reads "building @hellodrumbeat ! drumbeat ambassador" and nimaitalks' reads "side hustler making money with @hellodrumbeat". So the operation discloses on two accounts across both platforms, and on none of the others across either. The pattern is consistent, deliberate and platform-independent.

No paid-partnership label was present on any Instagram profile or post reviewed in this collection. An undisclosed, paid, briefed campaign about a political issue, delivered through accounts created for that campaign, engages all three Meta policies simultaneously.

Evidentiary note: what each platform exposes

Evidentiary note, stated precisely. The 6,048-post TikTok figure and the zero branded-content flags are from TikTok, where the platform exposes that field directly and every post could be checked programmatically. A separate TikTok field, isAd, is set on nine posts — all on the same pre-existing creator, mads_ventures. It is not the creator disclosure. isSponsored is false on all nine, none displays a paid-partnership label, and the nine include two charity appeals, a how-to about car pricing, a rant about people not reading captions, and a song cover — an incoherent set for an advertising flag. Reading the spoken-audio transcripts, however, two or three of them do carry an apparent brand relationship: one thanks @Milk-Bone “for making this moment possible”, and one tags @Kroger and @U by Kotex while promoting a named sale price. On the video frame that second post reads as a complaint about tampon prices; only the audio reveals it as a price promotion. None of them is labelled. This is noted because it is the same account, and the same absence of a label, as the political material — but no paid relationship is established here, and none is claimed. Instagram does not expose an equivalent field at collection scale, so the Instagram finding rests on profile-level and sampled post-level review rather than an exhaustive per-post audit. The distinction does not change the conclusion — but the two claims do not rest on identical evidence, and I am not going to pretend otherwise.

Two places where the money left a name

Across 8,463 persona posts these accounts almost never say who is paying them. Twice, the mask slips — once because a campaign committee required a label, and once because three accounts started campaigning for a United States Senate candidate and labelled nothing at all.

Two strands break the pattern this report describes. Both were found late, in the last collection pass, and both cut against earlier versions of this section. They are set out in full because they are the only places where the question everything else in this report cannot answer — who paid, and for what — is answered on the face of the posts.

One: a campaign committee, named by the creator

Nineteen posts on rae.luvs.ny — a New York persona, not a recruiter — carry TikTok's paid-political disclosure in full:

Three things follow. First, the client is named — a candidate committee, with its identification number, disclosed by the creator on every post in the run. Second, the personas are not tied to the state they claim: a persona whose bio reads "NYC girl with opinions" was paid to advocate in a California race, which is difficult to reconcile with the "sound local" premise the rest of the network runs on. Third, and most usefully, it proves these creators know how to disclose and that the mechanism exists on the platform. When a campaign committee is the payer, the label appears. Across the other 8,408 undisclosed persona posts, it does not.

What this does not mean. It does not follow that Steyer for Governor commissioned anything else in this report, or that the committee had any relationship with the operator of the wider network — nor that the operator brokered this run. The disclosure names the committee as payer. It does not name Drumbeat, and no intermediary appears on any of the nineteen posts. On the evidence it bought a short, properly-labelled run on one account. It is the one strand in this investigation that was disclosed correctly, and it is included for that reason as much as any other.

Two: a federal candidate, with no disclosure at all

Between 16 July and 5 August 2026, 79 posts across three Michigan accounts — lily.cather (36), bergen.ellis (29) and jennopalis (14) — advocate for Haley Stevens, a Democratic candidate for the United States Senate. Not in the oblique register the rest of the network uses. Directly:

Zero of the 79 carry any disclosure. 81,177 views. This is the cell that showed 30% partisan attribution in §04 — figures so far outside the rest of the network that they read as an error until you see what Michigan is actually posting. It is not economic grievance with a political tint. It is candidate advocacy — and it is running on accounts this report has otherwise documented posting about grocery prices.

Both exceptions in detail

This changes the legal reading in Annex C, and it changes it in the direction of more exposure, not less. That analysis concludes that federal disclaimer requirements are "largely not triggered" because the network avoids express advocacy. Michigan does not avoid it. "Why she is voting for Haley Stevens" is express advocacy for a clearly identified federal candidate, which is the precise trigger for 52 U.S.C. §30120. Whether a disclaimer was legally required turns on facts not visible from outside — who paid, how much, and whether the spending was coordinated or independent — but the content-side condition that the rest of this network carefully never meets, Michigan meets 79 times.

Who ran it is a separate question from who these accounts are, and the evidence answers only the second. All three Michigan accounts are network members on the same follow-graph test applied everywhere in this report — jennopalis is in fact the single best-connected account in the roster, with 29 links to other members, and bergen.ellis has seven. But membership does not establish that the operator of the affordability campaign commissioned the Stevens content. jennopalis is a pre-existing creator — opened September 2025, 257 videos, a real audience — and a creator like that can take a second brief from an entirely unrelated buyer. The defensible finding is that accounts in this network also ran undisclosed candidate advocacy, not that this network ran a Senate campaign. No payer is named on any of the 79 posts, and none was identified.

Per-account follow-graph evidence

Where each account stands, so the claim can be checked. jennopalis: 23 outbound and 6 inbound roster links, 14 of the posts. bergen.ellis: 6 outbound, 1 inbound, 29 posts. lily.cather: 2 inbound and nothing outbound — exactly the minimum this report requires, and it carries 36 of the 79 posts. That is the weakest link in this finding and the first thing a challenge should test; the finding survives without it on the remaining 43 posts, but it would be materially smaller.

Ten states, opposite tradecraft

Every node is the account's real profile photograph. Lines are verified follows between members. Switch platforms, recolor by cell or by provisioning wave, and drag any node to pull the layout apart.

The graph itself is in §01. This section is how it was built: which links were accepted, which accounts cleared the threshold, and which were rejected despite clearing it.

The Ohio cell interlinks heavily — @oliviainohio alone is followed by eighteen other members. The New York cell barely links at all: two internal edges and a single bridge, its.emma.in.ny lexigray181. Had the video contained only the two New York accounts, follow-graph analysis would have surfaced almost nothing.

The Georgia cell was found by following that logic. Seven Atlanta-anchored personas running the identical affordability script on the same wave structure, with two disclosed Drumbeat ambassadors of their own. One of them, jalex2792, states in its bio: "NOT affiliated with any political party." Nobody volunteers that unprompted.

Georgia turned out to be the largest cell, not a footnote. Expanding through the company's own Georgia account surfaced eleven further Atlanta personas, each carrying between four and eleven inbound follows from confirmed members, several announcing the brief in the bio itself — "Let's Talk Affordability in Georgia", "dreaming of an affordable georgia". Georgia is now the largest cell at twenty-six accounts, with Ohio just behind at twenty-five. One of the eleven, treysgeorgiathoughts, follows exactly one account on the entire platform: loladrumbeat, the paid ambassador. Full roster in Annex E.

The enumeration, account by account

The thin cells were then worked deliberately, because a one-account "cell" is not a cell. Pulling the follow lists of every member in New York, North Carolina, Pennsylvania and Maine and looking for co-located peers took North Carolina from three accounts to seven and New York from three to five. tyrants51 ("Charlotte, NC born and raised") was created 26 minutes after emma.smith.733 on 18 May — the same co-provisioning signature that exposed Maine, where two accounts were created eleven minutes apart. katie.saves20 ("Just an NC mama", five follows) posts "Is the American dream really dead? #northcarolina #costofliving"; derektalkss (NYC, five followers) posts "energy costs in New York City has been going up a lot and who is it to blame? The republicans!!"

Wisconsin is the seventh state, and it was found by handle alone. rantswithsaum ("Midwest Wisco") tags Alliant Energy — completing a pattern that runs three deep: Ohio personas attack FirstEnergy and AEP, North Carolina attacks Duke Energy, Wisconsin attacks Alliant. Same script, local utility swapped in. A second Wisconsin account follows it, and it is mutually linked with a Pittsburgh persona. The naming convention travels too — #ohiocheck, #wisconsincheck, and the same construction for every other state.

A provisioning wave four months earlier than the one this report was built on. Working the thin cells surfaced a cluster nobody was looking for: renee.ohio (19 Feb), addy.sam24 (19 Feb), rants.with.aimee (20 Feb), annacole5950 (23 Feb) and katie.saves20 (27 Feb) — five accounts across Ohio, Pennsylvania and North Carolina inside eight days in February 2026. Wave 0 in §03 begins in June. It is now clear that June was not the beginning.

Maine still did not substantiate, and that is reported as found. Maine remains a single account. It was searched directly and through its own follow list, and nothing else surfaced. Either those states are genuinely thin, or the method that works on dense cells does not reach them — the New York cell was nearly invisible to follow-graph analysis for exactly this reason, and it turned out to be real.

Two accounts cleared the numeric threshold and were still excluded. ncbudgetandtax has two inbound follows from confirmed members, but it is the account of a real North Carolina policy organization working on poverty and tax — network members following a think tank is expected, and counting it would be an error. pennsylvaniareals is more interesting: "PA's next-gen newsroom powered by local creators… by @Get Real News" — a different company running the same creator-newsroom model in Pennsylvania. It is not Drumbeat and is not counted here, but it is direct evidence for the caveat in §09 that competitors running this playbook would be invisible to a Drumbeat-anchored investigation.

Two further states came out of the sparse-cell logic. North Carolina — emma.smith.733 ("professional yapper", Charlotte) and hunterrrr972 ("Just yapping"), follow-linked to both Ohio and Georgia and running the Duke Energy equivalent of Ohio's FirstEnergy material. Pennsylvania — kailyn.moyer, reached from two different state cells. That took it to six states at the time. There was never a reason to think six was the number: New Hampshire was swept and came back empty, and a Michigan candidate matching every content and template test has no follow edge yet, so it is not counted here.

Not every account here was built for the job — and the exceptions matter. Three structurally different things are running under one campaign. Most are purpose-built personas, registered in the Jun–Aug 2026 waves with double-digit follower counts. Two are reactivated dormant accounts — kelsie.explains (2021) and ganewswithparis (2024) — invisible to creation-date analysis. And at least one is an established creator using their own profile: mads_ventures, registered October 2019, 11,100 followers, 410 videos, a UGC business address in the bio and a pinned playlist titled "georgia affordability". That last category is the most consequential for disclosure law — it is the only one where a real audience, built over six years for other reasons, is being served briefed political content without a label.

The most-followed nodes are the paid ones. Sort the graph by inbound follows and the top two are loladrumbeat (21) and marya2cents (20) — both publicly identifying as Drumbeat ambassadors, ahead of every undisclosed persona. The disclosed layer sits at the center of this graph, not at its edge.

Where the workings are. How every figure in this report was built, how each account was tested before being counted, the full legal reading, and the leads that were tested and closed are set out in the technical annex below. Nothing in the body depends on a step that is not written down there.

Two accounts do not independently write the same sentence

Two different people. Two different accounts. One biography, character for character.

One hundred and thirteen accounts is not one hundred and thirteen people

A caution that cuts against the headline number, and it belongs here rather than in a footnote. Several confirmed members are the same operator running more than one persona — four handle families in all, one of them three accounts deep. The clearest is a pair that publishes the same contact address on both profiles, toolie.xo@gmail.com. The account count is a measure of deployed infrastructure, not of headcount. The number of human beings behind this network is smaller than one hundred and thirteen, and this report does not attempt to say by how much.

That cuts both ways, and the second direction is the one that matters for enforcement. One person operating four labeled personas is not four creators making an honest mistake about disclosure — it is a single actor maintaining four separate fictions at once.

Left: rants.with.kay, 91 followers. Right: rantswit.breezy, 0 followers, created twelve days later. Same bio, same emoji, same parenthetical.

The templates underneath

Six writing formulas recur across otherwise unrelated handles.

The bio and handle evidence

The "born and raised" formula is the only linguistic thread crossing from the Ohio cell into New York — the follow graph separates those two almost completely, but the persona writing does not.

The expansion made this formula more visible, not less. "Born and raised" (or "born and bred") now appears in ten biographies across seven states, each swapping in its own city — Manhattan, Detroit, Austin, Charlotte. The localization is real; the sentence holding it is not. The Texas cell adds a second frame of its own, three accounts opening "just yapping about" and varying only the object: the world, whatever, Texas.

And one pair that is almost certainly one person. talkswithharv reads "talking about the things that matter! Austin, Texas"; harvtalks reads "talking about the things that matter! Austin, Texas born and raised". Same opening sentence, same city, same first name inverted across the handle, created eight days apart. talkswithharv follows harvtalks. This is the clearest instance in the roster of one operator running two accounts, and it is the reason the headcount caveat above matters.

What this does not prove. "Born and raised" and "Mom of two" are ordinary English, and neither frame is evidence on its own. No account is placed in this network on bio wording alone — every one also had to clear the follow-graph threshold. The templates corroborate; they do not carry the claim.

The same handle, the same content, twice

Every one of the 113 handles was checked at the identical address on instagram.com/<same username> — no third-party identity tool required. Where no account exists, that negative is recorded rather than left blank, so the rate below has a complete denominator. And because a matching handle is not by itself a matching person, each account found was then tested against its TikTok twin on bio text, post count and profile photograph before being called a mirror. Then the two profiles were compared directly: bio text, profile photograph, and post count.

Correction: why the mirroring rate fell

Correction, and why the number went down. An earlier version of this report put the mirroring rate at 83%. That divided by the handles checked at the time (49 of 59), not by the whole roster. Checking all 113 — and recording the negatives — gives 73%. The rate fell because the denominator got honest, not because the network changed.

The follow list is not social. It is targeting.

These accounts follow a lot of local media, restaurants, realtors and neighborhood pages, and the obvious reading is camouflage — dressing the profile so a human visitor believes the persona lives there. That reading is probably too small. The operator says what the follows are actually for, in its own pitch:

Following local accounts is one of the clearest locality signals available to a recommender system. On that reading the follow graph is not a social artifact at all — it is an input to the targeting, doing the same job as the state in the bio and the local utility in the script. The persona follows Ohio accounts so the algorithm files it as Ohio and serves it to Ohioans.

The follow data supports this. Comparing the external follows — everything outside the network itself — accounts in the same cell overlap far more than accounts in different cells:

Same-cell accounts share outside follows at twenty-three times the cross-cell rate, and that is on five times fewer pairs. Ohio personas follow the same Ohio accounts as each other; Georgia personas follow the same Georgia accounts. What that describes is a starter list issued per cell — not twenty-two people who independently discovered the same neighborhood pages.

And the lists are mostly not news. The most-shared outside follows in the Ohio cell are ordinary local personal accounts — a campus-life page, several Columbus individuals — carried by nine to twelve personas each. They are named nowhere in this report, because they appear to be uninvolved local people whose only connection is being on someone's list. The pattern matters; their identities do not.

A lead tested and retired

This retires a lead, and it is worth saying so. One outlet, couriergeorgia, is followed by eight roster accounts — more than any other news outlet in the network, where nothing else exceeds two. Taken alone that looked like a relationship worth chasing. Measured properly it is not: within its own cell it is tied with an unrelated local Atlanta account, both on seven; the Texas and North Carolina cells follow their equivalent outlet zero times across nineteen accounts; and it is cited as a source in four posts out of 9,261. It is one entry on Georgia's locality list. No inference about funding, ownership or direction is drawn from it, and none is supported.

Why some cells are harder to enumerate

Two shapes of account, and it changes how findable they are. Sweeping the Instagram graph out from members across the state cells showed the cells are not built alike. Some accounts follow almost nothing but each other — oliviainohio follows ten accounts and five are network members or the operator itself. Others follow almost nothing but local camouflage: kailyn.moyer's 26 follows are Pittsburgh news, realtors and food pages with no network members at all, and annacole5950's are Raleigh tourism and business accounts. An investigator who happens to start on a camouflage-heavy account finds nothing and concludes the cell is small. Starting one account over surfaces half the network. The apparent size of this operation depends heavily on where you begin, which is a further reason to read 113 as a floor.

Why the method matters more than the correction

Why the method matters more than the correction. A follow graph and a handle-match test answer different questions, and the second one inherits an assumption the first does not. Checking instagram.com/<same username> can only ever find mirrors that chose the same username; it is structurally blind to the ones that did not, and it returns a confident negative either way. That is the most dangerous kind of measurement error — not a missing number, but a wrong one that looks complete.

What the posting clock rules out

Which corpus this section uses

Which corpus this is. This section plots the 6,048 TikTok posts for which an exact publication timestamp could be recovered — the whole TikTok catalog, not the 542-post sample an earlier version of this section relied on. Instagram is deliberately excluded here: a mirrored post is the same human action seconds later, and counting it twice would distort both the hour profile and the gaps between posts.

Posting volume over time, one row per cell. Taller and darker means more posts in that week.

Density, not synchronization, is the signal here. Accounts sustain one to three posts a day for weeks — the published quota is 30 posts a month — but they do not fire in unison.

When they post

All 6,048 TikTok posts by hour of day, US Eastern.

93% of all posts land between 08:00 and midnight Eastern, on a broad plateau that runs from late morning to midnight rather than a single evening spike. This is a finding that cuts against a foreign-operation reading: the network keeps American waking hours, with the heaviest output after the American working day. It is consistent with real US-based people posting around a job — which is exactly what a paid domestic creator program would produce.

The timing analysis

I also tested for batch uploading and largely did not find it: only 4.0% of consecutive posts are less than ten minutes apart, now measured across 5,942 intervals rather than a few hundred. The figure did not move when the corpus grew elevenfold. Three accounts stand out — lily.cather (54% of its intervals), rants.withgloria (38%) and annacole5950 (29%) — and whocanaffordthis publishes in pairs 77 to 166 seconds apart before going quiet for twenty hours. But it is not a network-wide behavior, and I am not presenting it as one.

Real people, new accounts, a supplied script — and no disclosure

Every structural finding below is directly verified from platform data. The commercial mechanism is documented from the operator's own published material. The one thing I do not resolve is which client paid for it.

Hover or tap any finding for the evidence behind it.

• The four accounts in the video were batch-created. Proven · High + details Platform creation timestamps: 30 Jul 08:09, 30 Jul 22:20, 30 Jul 23:46, 31 Jul 18:41 UTC — a 34.5-hour span. Confirmed independently by two collection sources agreeing to the second.
• Three pairs of accounts carry byte-identical biographies. Proven · High + details Browns fan (unfortunately) 🏈 appears on both @rants.with.kay and @rantswit.breezy. the affordability crisis is real and alive, especially in Ohio! let's chat — 74 characters — appears on both @ohioinfo and @newswithchloe. And Mama of 2 from the Bronx🌺 appears on both rants.with.ki and rants.withki — handles separated by a single full stop, created six days apart.
• Twenty-five of twenty-seven handles are mirrored on Instagram. Proven · High + details Twenty of those publish a post count within one of their TikTok count; eleven match exactly. That is one content stream pushed to two platforms, not two accounts run in parallel.
• One persona's Instagram name still contains its staging note. Proven · High + details @rants.with.rae's Instagram display name reads Rae-not active yet- while its TikTok side runs 22 posts to 39 followers. Nobody arrives at that name through personal use.
• The follow graph splits into ten state cells. Proven · High + details 799 internal follow edges across 113 accounts in ten states; the two most-followed nodes in the whole graph are disclosed paid ambassadors. The New York personas instead follow real local food and lifestyle creators — better camouflage, and far harder to detect.
• Accounts in the same graph disclose that they are paid. Proven · High + details @kamyapss and @nimaiyaps state "building @Drumbeat" / "DrumBeat Ambassador" in bio; @kamyapss recruits publicly — "if you want to get paid to talk about local Ohio issues like this, comment down below LOCAL." @loladrumbeat runs the Georgia equivalent.
• The published method explains every anomaly here — and the paid ambassadors are wired into the network. Probable · High + details Drumbeat Labs LLC publishes its process: creators "establish new TikTok and Instagram accounts and post 2-3 times daily with provided content inspiration," $500 per 30 posts, with affordability among its named campaign verticals. Eleven separate observations follow from that one paragraph. Beyond method resemblance, the four self-declared paid ambassadors are directly follow-linked to 12 of the undisclosed personas across 28 edges, including 7 mutual pairs, and one of them follows the company's own TikTok account. A competitor merely copying the playbook would not be wired in like that. Still assessed rather than proven: no contract or payment record was obtained. See the attribution chain in §02.
• Persona accounts disclose on 55 of 8,463 posts — 0.65%. Proven · High + details All 6,048 TikTok posts return the platform's branded-content flag as false; no #ad or sponsorship marker appears in any caption or persona bio; and no paid-partnership label was found on any of the 25 mirrored Instagram accounts. This breaches TikTok's Branded Content Policy — which also bans political branded content outright — and Meta's mandatory paid-partnership labeling.
• It breaks both platforms' rules, and may break election law. Proven · platform rules Open · statute + details Undisclosed paid posting breaches TikTok's Branded Content Policy — which bans political branded content outright — and Meta's mandatory paid-partnership labeling. Whether it also breaches Ohio and New York disclaimer statutes depends on whether the paying client is a political committee or a nonprofit, which is not established. See Annex C.
• What is documented here is a fragment of the operation. Proven · operator's own figures + details Drumbeat's founders state publicly that the company has generated "70M+ cost-efficient views… across six pilots" in nine months, plus 2.3 million views for a single nonprofit partner in under seven days from 70+ posts. This report maps ten state cells and 9,261 posts. Every enumeration method used here also has a known blind spot — see §09. The network is larger than 113; how much larger is not established.
• Who funded the campaign. Open + details Drumbeat says opportunities come from "partner organizations." None are named publicly, and no first-party contract, brief or payment record was obtained. This single unknown is what decides whether the campaign is merely against the rules or actually unlawful.

The accounts, their timing, their duplicated assets, their mirroring and the disclosed recruiter layer are verified. The link from these personas to that specific platform is assessed. The paying client stays open.

The technical annex

Everything above is the finding. Everything below is how it was established: what was collected, what the thresholds were, what was tested and rejected, and where this account of the network is weakest. It is here rather than in the body because a reader deserves the story first and the workings on request — but the workings are the reason the story stands up, and nothing in them is hidden.

How the reach figures were built

The 63.4 million figure counts campaign content only. This is what was excluded, and what the number cannot bear.

What counts, and what was thrown out

Fourteen of these accounts existed before the campaign and have back catalogs of ordinary personal content. Counting that would inflate the number with videos that have nothing to do with the operation. So every post was classified on what is actually in it — machine-read scene descriptions capturing on-screen text, plus the caption — and only posts about affordability, rent, groceries, utilities, wages, taxes, healthcare, childcare, student debt, data centers or public spending were counted.

Only 33 posts in the entire catalog cannot be read at all — no caption, no on-screen text, no description — and between them they carry 1,871 views. An earlier version of this report had to exclude 546 posts and 4.7 million views on those grounds; reading the video frames rather than the captions removed almost all of that doubt. The 33 are excluded, which makes 63.4M a floor rather than an estimate, but it is now a floor with very little air beneath it.

Reading the on-screen text, and why the platform’s own text could not be used

These are talking-head videos with the argument written across the frame. Any measure of what the network says therefore depends on reading that text, and the obvious source for it turned out to be unusable.

TikTok exposes a machine-written scene description for some videos. It is good — it renders on-screen text in about four of every five videos where it exists — but it exists as a function of recency, not content. Coverage across this corpus runs at zero percent for every month before January 2026 and climbs to 86 percent by August 2026. The same pattern appears account by account: some accounts carry it on every post, others on none, according to how a collection run happened to be ordered. TikTok’s subtitle and transcription fields behave the same way and return empty for all 2025 posts, in profile collection and in single-video collection alike.

An instrument like that cannot be used to compare anything. A theme measured with it would appear to grow through 2026 for the sole reason that the instrument improved, and an account would appear more on-message than its neighbor because its posts were read more thoroughly. Cross-state, cross-account and month-on-month comparisons built on it would be measuring collection, not behavior.

What is used instead is optical character recognition of each video’s cover frame. This is indifferent to a post’s age: it applies the identical operation to a September 2025 post and an August 2026 post, and it returns the words as written rather than a paraphrase. Where the machine description said a video discussed surveillance, the frame itself reads “flock is evil”. It runs on 99.9 percent of the TikTok catalog — every post holding a retrievable cover image.

What that still does not reach. The cover is one frame. Where a video’s text changes as it plays, this captures the opening card and not the rest, so it reads the subject reliably and the full argument only sometimes. Instagram publishes no equivalent field and no cover was retained for its posts; because the two platforms carry the same videos, an Instagram post inherits the reading from its TikTok twin where the pair can be matched on handle, caption and date, and stays caption-only where it cannot. Both limits push the same way: they can only cause campaign content to be missed, never invented. Every figure here is a floor.

A merge fault, found and corrected. Posts collected in more than one pass were being reconciled by choosing whichever record looked richest, rather than by combining their fields. Where an early pass had returned a post with its caption stripped, that blank record could survive a later pass that carried the caption intact. The fault concealed the text of several hundred posts, among them one carrying 1.6 million views and the caption “they don’t want us to talk about this #costoflivingcrisis” — unambiguously on-topic, and counted as unreadable. Records are now combined field by field, and the posts that remain genuinely textless number in the tens rather than the hundreds.

What these numbers are not

They are not selling anything either

One test worth running against a paid network is whether the accounts monetize the audience they are building. They do not. All 9,261 collected posts — 6,048 on TikTok, 3,213 on Instagram — were searched for the ordinary machinery of creator commerce. The result is a clean zero.

This sharpens the central observation rather than softening it. The purpose-built personas have audiences, they post daily, and across 5,910 campaign posts they never try to sell anything — no code, no link, no product. (The pre-existing creators do carry ordinary brand work: nine posts in the corpus are flagged consumer deals, all on one account. Not one is political.) A creator building a following for their own benefit monetizes it; that is the entire economic logic of being a creator. These do not, because the audience is not the asset being sold. It has already been sold, to whoever commissioned the message. The absence of a commercial ask sits beside the absence of a political one, and both point the same way.

Which Instagram number this uses, and why it is the smaller one. Instagram reports two figures per video. videoPlayCount counts plays including autoplay and loops; videoViewCount counts views. On the same posts in this corpus the play figure runs 2.3 to 2.9 times higher — the single largest post reports 9,033,546 plays against 3,215,486 views. Every Instagram figure in this report uses the view count, the smaller and more TikTok-comparable of the two. Using plays instead would put the Instagram total at 44.6 million rather than 17.1 million, and the headline at roughly 88 million. That number would be defensible on Instagram's own labelling and is not used here.

A collection gap, corrected. Until 17 August the Instagram figures rested on 44 of the 88 confirmed mirrors — the others had been identified but never had their posts pulled, so their reach sat outside every total. Collecting the remaining 44 added 2,015 posts and moved the campaign figure from 30.4 million to 43.3 million. Reading the on-screen text then moved it again, to 63.4 million. The first correction is a reminder that a mirror confirmed is not a mirror measured; the second, that a post collected is not a post read.

A gap in the keyword set, found and closed. The anchor list was audited against the posts it rejected. It was missing renters, bare daycare and the cost of… construction — core cost-of-living vocabulary rather than generic price words. Adding them moved the campaign figure by 0.1%, fifteen posts and 78,217 views, every one of them plainly on-brief. Bare economy was tested at the same time and deliberately left out: it pulls in a post about media bias and one about American exceptionalism, neither of which is affordability content. The rule stays narrow on purpose — an earlier, looser version of it inflated the total by roughly sixty percent.

How accounts were tested before being counted

Every account here had to clear a follow-graph threshold, and every cross-platform mirror had to clear a three-attribute identity test. Both are set out here, including the cases that failed them.

A matching handle is not a matching person

Usernames collide. Every one of the 82 accounts found was therefore compared against its TikTok twin on three independent attributes — bio-text similarity, post-count proximity, and a perceptual hash of the profile photograph. Two or more attributes agreeing is treated as confirmation; one is probable; none is a handle and nothing else.

That last row is the reason this test exists. Four accounts would have been counted as mirrors on handle alone, and all four are something else — most likely the real person behind an ambassador account, using Instagram as themselves. Counting them would have inflated the mirroring rate and, worse, implied campaign content on accounts that are not carrying any.

Three rates, and they are not the same number. 78% of the roster (88 of 113) has an Instagram account. 60% (68 of 113) is corroborated as the same persona on at least one attribute. 50% (56 of 113) is confirmed on two or more. An earlier version of this report quoted only the first of those and called it a mirroring rate; the honest headline is the third. The difference between 73% and 44% is the cost of checking, and it runs in the direction that makes the network look smaller, not larger.

Six mirrors are claimed but empty — and that is its own finding. Someone registered the name and never posted to it. It is the signature of provisioning ahead of need: the handle is reserved so the persona can be pushed to a second platform later, and for six accounts that step has not happened yet.

Mirroring is not uniform, and Texas is the outlier. Ohio has an Instagram account at 88% of its handles (24 of 27) and the ambassadors at 83%. The Texas cell — the newest and fastest-built — sits at 53% (8 of 15). The cells that have existed longest are the most completely deployed, consistent with the second platform being a later step in a standing process rather than part of the initial build.

The operator mirrors too, and says the quiet part out loud. instagram.com/hellodrumbeat exists and states the business in one line: "Connecting creators with paid opportunities to talk about stories, news, and issues they care about." The company describes the arrangement plainly on its own account. None of the personas do.

Ten accounts match on all three at once — identical bio, identical photograph, and post counts within two: paigerants3, rants.with.kay, its.emma.in.ny, hunterrrr972, oliviainohio, rantswithivy, talia.smith333, kelsie.explains, karonde.yaps and lexigray181. That is not a person who happens to use two apps. That is one asset, deployed to two platforms.

How the photo test works, and what it does not prove. Each pair was compared with a perceptual difference hash — a measure of whether two images are the same picture, robust to resizing and re-compression but not to a different photo of the same person. A score of 0 means the identical file. The 37% figure is therefore a floor on visual mirroring, not a ceiling: accounts scoring "different" often still show the same individual in another shot, which this test deliberately will not count.

Nineteen of the twenty-five mirrors sit under forty followers. @jazzmyneohio has published 54 Instagram posts to an audience of seven. Sustained daily publishing to nobody is only rational if payment is per post rather than per result — which is the compensation model the platform publishes.

The handle is not always the same handle

Everything above assumed the mirror sits at the identical username, because that is what the network mostly does. Walking the Instagram follow graph out from confirmed members — a different route to the same question — showed the assumption is wrong often enough to matter. Six accounts recorded above as having no mirror, or an empty one, in fact have an active mirror under a slightly different handle.

This cuts one way only. Every correction found by this method adds a mirror that same-handle checking had scored as absent. The published rates are therefore floors: the true mirroring rate is higher than 78%, and the true verified rate higher than 50%. How much higher is not established, because exhaustively testing handle variants for all 113 accounts has not been done.

Two accounts that are not cells

Maine and Iowa appear in every table in this report as though they were cells like the others. They are not. Each is a single account, and neither is held in place by the follow graph that defines every other cell here. They are counted, but they rest on a different kind of evidence and a reader should know which.

An eleven-minute creation gap is not weaker evidence than two follows — it is different evidence. Accounts can follow each other for many reasons; being registered eleven minutes apart on the same afternoon, and then posting the same kind of content in a different state, is a coordination signal that is hard to produce by accident. It is also the reason this account was found at all: it has no graph trace, and nothing but the timestamp connected it to anything.

And there is no cell behind either of them. A hashtag sweep of Iowa and Maine cost-of-living content since April 2026 returned 100 posts and not one further persona account — the rest is sorority recruitment, the Des Moines Register, local photographers, state-fair coverage. Every other state in this report has between four and twenty-eight interlinked accounts. These two states have one apiece and no internal structure, so "the Maine cell" and "the Iowa cell" overstate what was found. Read them as two isolated accounts that the evidence places inside a network built elsewhere.

What would settle it. For Maine, any follow edge in either direction, or a second Maine account in the same provisioning window. For Iowa, one more inbound follow. Both accounts are live and both are worth re-testing as more of the network's follow graph is collected — 27 of the 113 accounts still have none.

The legal reading, and its limits

Platform rules, FTC endorsement principles, federal and state election law, and the coordinated-inauthentic-behavior framing — what each reaches, and what it does not.

The FTC principle

The FTC's Endorsement Guides (16 CFR Part 255, revised 2023) rest on a single idea: where an endorser has a material connection to whoever is behind the message — money, free product, employment — the audience has to be told, clearly and unmistakably, in the post itself. Payment of $500 per thirty posts is a material connection by any reading. Judged against that principle, this network fails completely.

Whether it also constitutes a legal violation is a narrower question, and I am careful about it here for a reason. The FTC Act reaches deceptive practices "in or affecting commerce," and its jurisdiction is over commercial speech; pure political or issue advocacy has historically sat outside it. If the paying client is a nonprofit running issue advocacy, the Guides are a poor jurisdictional fit even though the conduct is exactly what they exist to prevent. That gap is not a technicality to be waved away — it is the finding. The operator's own marketing calls the product "like geo-targeted advertising" and claims it beats "paid advertising" on trust, while the arrangement is structured so that neither advertising disclosure rules nor election disclaimer rules clearly bite.

Election law: where the exposure actually sits

Undisclosed paid political communication is regulated in the United States — but the rules bind the payer, and they turn on who that payer is. That is the single fact this investigation could not establish, and it is why the report stops short of calling the campaign illegal. Set out plainly, so the test is visible:

Campaign finance: the internet exemption, and the one route around it

Campaign-finance law is where most readers expect this to land, and it is worth being precise about why it mostly does not — because the reason is itself the story.

Federal reporting rules attach to a "public communication." Under 11 CFR 100.26 that term covers broadcast, cable, satellite, newspapers, mass mailing, phone banks and paid internet placement. Organic social posts are specifically outside it. A creator paid $500 to post on their own account is not buying ad space; nothing is "placed for a fee on another person's website." The same logic exempts this content from electioneering communication rules under 52 U.S.C. §30104(f), which reach broadcast, cable and satellite only. Independent-expenditure reporting under §30104(c) requires express advocacy — "vote against," "defeat" — and across 9,261 posts this network almost never uses it.

So the money moves without a filing. Not through a loophole anyone had to find, but through the ordinary shape of the rules: pay a person rather than a platform, use grievance rather than express advocacy, and the federal disclosure architecture does not engage at all.

The exception is coordination, and it is the sharpest unresolved question here. Under 11 CFR 109.21, spending made in cooperation with, or at the request or suggestion of, a candidate or party committee becomes an in-kind contribution — reportable, subject to limits, and unlawful if the source is a corporation or over the cap. The content prong of that test can be met by a communication that refers to a clearly identified candidate within the pre-election window. This network does refer to clearly identified candidates: posts naming Jon Husted, and one explicitly framing "Brown's record vs. Husted's record — the Ohio Senate race." If any of that was produced in coordination with a campaign or party committee, the analysis changes completely.

I have no evidence of coordination and am not alleging it. I am identifying the test, because it is the one route by which this conduct becomes a reportable contribution rather than an unregulated purchase — and it is answerable only by someone who can see contracts and payments. Note also that the campaign now spans ten states, each with its own disclaimer and electioneering regime; the conditional in the table below multiplies accordingly, and Ohio and New York are simply the two examined here.

So the answer to "is this illegal" is not a shrug, it is a conditional with one unknown in it: if the organization that paid for this campaign is a political committee, then a coordinated program of paid political communications carrying no disclaimer, on two platforms, across 8,408 undisclosed persona posts, is a serious problem under state election law. If the payer is a nonprofit doing issue advocacy, it likely sits outside those statutes — and outside the FTC's commercial-speech jurisdiction too, while still breaching both platforms' rules.

That is the finding, and it should not be softened: the conduct is identical either way, the deception of the audience is identical either way, and whether it is unlawful depends entirely on a fact the public cannot see. Drumbeat names no clients. The personas name no sponsor. A regulator with subpoena power could resolve this in an afternoon; a member of the public in Columbus cannot resolve it at all.

This conduct is already regulated in one of these ten states

The disclosure analysis above is not hypothetical, and it is not only federal. In June 2024 the Texas Ethics Commission voted 7–0 to require anyone paid more than one hundred dollars to post or repost political advertising to say so. That rule was in force before a single account in this network began publishing, and Texas is the second-largest cell here — fourteen accounts and roughly 18.3 million views.

The rule was written in response to a named company. Reporting at the time identified Influenceable LLC, which had been paid to recruit social-media creators to defend the state attorney general during his impeachment. Recruit creators, supply the line, do not label the result. A regulator has already looked at that structure once and written a rule about it.

What that does and does not establish. The Texas rule reaches political advertising — communications supporting or opposing a candidate or ballot measure. The overwhelming majority of this network’s output is issue content with no express advocacy, and on the plain reading it sits outside that definition. That is not a gap in the finding; it is the finding. The content is pitched precisely where the grievance lands and the disclosure duty does not. Where these accounts did cross into candidate advocacy — the seventy-nine undisclosed posts discussed in §06 — the analysis is different, and that is the exposure worth watching.

And the law is moving toward this conduct, not away from it. In August 2026 California was considering a bill from Assemblymember Marc Berman allowing the Fair Political Practices Commission to fine creators and committees up to $5,000 per violation, and — the operative part — to do so administratively rather than through a court process that can take months. New York’s legislature has weighed a comparable requirement. California matters here for a second reason: a California gubernatorial committee is one of only two payers this network ever named, on the nineteen disclosed posts in §06. This report takes no position on any pending bill. The point is narrower and factual: the practice documented here is already governed in one of the states where it ran, is under active legislation in others, and has a named regulatory precedent built the same way.

Call it what it is: coordinated inauthentic behavior

The industry term for this is not "influencer marketing." It is coordinated inauthentic behavior — and this network meets the definition on every element, not loosely but precisely.

The obvious objection is that the people are real — real faces, real voices, no bot farm, no stolen photographs at the network level. That objection is worth taking seriously, and it is why this report says plainly in §09 that this is not a bot network. But it does not rescue the conduct, for two reasons.

First, the inauthenticity is in the account, not the human. "Kay, Browns fan, struggling with Ohio prices" is a constructed vehicle: created for the campaign, scripted, mirrored to a second platform, and — decisively — carrying a biography copied word for word onto another account. A persona whose bio is duplicated across two identities is not a person expressing themselves. It is a template being filled.

Second, the precedent is settled. Meta has repeatedly removed networks it classified as coordinated inauthentic behavior that were run not by states but by commercial marketing and PR firms on behalf of paying clients, staffed by real people. "There was a real human behind the account" has never been the test. The test is whether the audience is deceived about who is behind the coordinated activity and why — and here, across 8,408 undisclosed persona posts on two platforms, they are told nothing at all.

What distinguishes this case from the state-actor archetype is not authenticity. It is that the operation is domestic, commercial, invoiced, and advertised on LinkedIn. That should make it easier to address, not harder.

I am an intelligence analyst, not a lawyer, and none of the above is legal advice. The platform-policy finding is stated as fact because it follows from published rules; the statutory questions are stated as questions because their answers depend on the client's identity and legal character, which is not established.

Attribution, limits and retired leads

How far the attribution goes, what is deliberately not claimed, and the leads that were tested and closed.

How this is attributed — and how far it goes

Attribution is the part of any investigation most often asserted and least often shown. Here is the actual chain, each link graded.

Link 4 is what moved this case. Until it was collected, the argument rested on method resemblance — and a competitor copying a published playbook would produce the same observables. It no longer rests there. Half the undisclosed network is one follow-hop from someone who says in their own bio that Drumbeat pays them, in both directions, with seven mutual pairs. A rival firm's creators would not be wired into Drumbeat's declared ambassadors like that.

Note also that marya2cents — who runs the Georgia campaign — follows six of the Ohio personas. Genuine locals in Savannah do not follow six obscure Columbus affordability accounts. Creators enrolled on the same platform do. The same cross-state wiring runs the other way: the Ohio account kelsie.explains follows the Georgia persona paigerants3.

And it is not an isolated post — it is a recruitment pipeline running since at least March. Searching the #drumbeatambassador tag surfaces two earlier examples, both predating the affordability campaign entirely:

"Get paid based on your views" is the whole model in six words — performance-priced political content, on accounts that carry no label. And the March post proves the machine was recruiting five months before the Ohio personas were provisioned, in a vertical (psychology) with nothing to do with affordability.

The company account itself. @hellodrumbeat — created April 2023, 1,298 followers, 42 posts, bio: "Connecting creators with paid opportunities to talk about stories, news, and issues they care about." It follows 737 accounts.

A boundary I am not going to cross. That follow list is dominated by established political creators — several in the 40,000-to-500,000 follower range, including one focused on Florida primaries and one sitting elected official. It is tempting to read it as a client roster and add 250 names to this report. I am not doing that. A company account following a creator is one outbound edge; it evidences interest or prospecting, not payment. Every account counted in this report cleared a different bar — two or more inbound follows from confirmed members, matching content, or an explicit self-disclosure. The macro tier is logged here as a lead for anyone with subpoena power or a creator willing to talk, and nothing more.

That is the recruitment step, captured in the act. It closes the loop the rest of this report infers: a Drumbeat ambassador account, created days before collection, existing for the sole purpose of recruiting more people to post local Ohio news for money. The eighth disclosed ambassador found.

Two accounts, one mailbox. toolieganews — the 126-post Georgia news persona — lists toolie.xo@gmail.com as its contact. So does mayor.tooliani, a second Atlanta account whose own posts are about the business model: "UGC changed my way of thinking about money and work… #sidehustle #ugccreator." The same operator runs the local-news persona and, separately, talks openly about UGC as income. Shared contact infrastructure is the hardest link in this report that does not rely on the follow graph at all.

One persona names its own operator. The Instagram bio of ganewswithparis — a Georgia account with eleven inbound follows from confirmed members — ends with the line "Main acc:" followed by a separate personal handle. The persona is explicitly labeled by its own operator as the secondary account. I am not publishing the personal handle: the individuals here are paid participants in a lawful program, not its architects. The structural point stands without it.

A company-branded account sits inside the graph. peachbrief — display name "Drumbeat Data", no bio, 26 followers, 7 posts — is followed by seven confirmed Georgia members and posts the same Atlanta affordability material as the personas around it. It is the third Drumbeat-named asset found, after @hellodrumbeat and loladrumbeat.

One further observation, offered as a lead rather than a finding: the same ambassadors also follow accounts running entirely different campaigns — a non-toxic-home products persona and an Atlanta nightlife "ambassador" — which suggests the affordability operation is one client among several on the same roster.

What that still does not prove. Following someone is not a contract. Creators in a niche do follow one another, and I hold this at Probable rather than Proven because no first-party record was obtained. The specific four accounts in the referral video are not individually follow-linked to an ambassador; they are placed in the network by the evidence in §07–§08 and in the campaign by the method match. And the client — the organization whose message this is — remains unnamed.

What I don't claim

Stated plainly, because the gaps matter as much as the findings.

• This is not a bot network. + details Profile photographs were tested for synthetic generation and came back negative — these are photographs of real people. The accounts are new, briefed and undisclosed; they are not automated.
• This is not a foreign influence operation. + details Nothing observed points overseas. The posting-hour distribution points the other way. An earlier hypothesis based on foreign-locale audio labels was tested and refuted — those were mass-trending sounds with tens of millions of plays.
• The named platform is a lawful business operating in the open. + details Drumbeat Labs LLC publishes its method, its rates and its recruitment. Nothing here alleges criminality. The finding is narrower: the method it publishes produces content that is indistinguishable from astroturf at the point a viewer sees it, and its public terms impose no disclosure obligation.
• The individual creators are not the story. + details On this evidence they are paid participants in a legal program. They are identified only by public handle. One persona photograph returns high-confidence face matches to a person posting under a different name; whether that is a creator using a pseudonym or an uninvolved person whose images were taken is unresolved, and that individual is deliberately not named.
• Thirty-four is a floor, not a total. + details Enumeration leaned on batch creation dates, and at least one confirmed member (kelsie.explains) is a dormant 2021 account reactivated in August 2026 — invisible to that method. Ten state cells were found; there is no basis for assuming ten is the number.
• One earlier judgement in this report was reversed. + details nnicoletalks was initially held as organic — it pre-dates the provisioning waves by eight months, has 920 followers and 101 videos, and covers weather and local news alongside affordability. On the expanded graph it carries 18 inbound follows from confirmed members, more than any other account outside the roster, posts the same affordability and Flock-camera material as the Ohio cell, and its bio advertises a media business and the role "social media manager". The category that resolves it — an established creator taking the brief on an account they already had — did not exist in my model when the first call was made. It does now, so the account is reclassified as probable. The original reasoning is left here rather than deleted.
• No effect was measured. + details Per-post view counts run from the hundreds to low thousands, with two outliers near 69,000 and 8,400. Capacity is not impact, and the instruments used here cannot measure whether anyone's belief or behavior changed.
• The referral compilation is a lead, not evidence. + details The video at the top of this page is someone else's selection of accounts. It is what started the investigation; every claim below it was re-collected independently from platform data.

Collected 15 August 2026 from publicly visible profile and post data. Drumbeat Labs was not contacted for comment before publication; the company’s own published account of its methods is quoted throughout instead.

The account register, and the captures

Ranked by how much each image proves without accompanying explanation.

The account register

Every account in the network, with the evidence attached to it. The narrative above names only the handles that carry a specific point; this is the complete list, so any claim about the size or shape of the network can be checked against it. Campaign posts and campaign views are that account’s share of the 7,519 posts and 63.4 million views counted in this report. Created is decoded from the account’s own numeric id. Eight of the 113 returned no retrievable posts at collection time and carry zeros.

Show all 113 accounts

Handle	Cell	Created	Followers	Campaign posts	Campaign views	IG mirror
earthluvr.444	Operator	2025-10-30	428	91	2,260,664	—
hellodrumbeat	Operator	2023-04-14	1,298	5	2,173	yes
kamrynyaps	Operator	2026-08-09	0	1	1	—
kamyaps	Operator	2026-07-01	5	2	24	—
kamyapss	Operator	2026-07-14	16	18	3,184	yes
loladrumbeat	Operator	2026-05-08	917	103	1,004,325	yes
luckyliberal	Operator	2026-01-12	731	8	8,602	—
marya2cents	Operator	2026-06-25	162	14	39,067	—
nimaitalks	Operator	2026-07-05	7	0	0	—
nimaiyaps	Operator	2026-06-23	68	29	74,473	yes
peachbrief	Operator	2026-07-23	26	7	9,326	—
rootedbean	Operator	2025-12-01	237	6	4,772	—
all.things.atl	Georgia	2026-08-05	44	18	14,889	—
benq404	Georgia	2026-07-20	38	53	52,253	yes
ganewswithlinds0	Georgia	2026-07-18	77	27	86,270	—
ganewswithparis	Georgia	2024-01-12	783	160	868,470	yes
henkeisjj	Georgia	2025-11-06	1,529	219	14,289,007	yes
itssemi.ga	Georgia	2026-07-18	49	43	71,664	—
jah.in.ga	Georgia	2026-07-30	30	12	11,304	—
jalex2792	Georgia	2026-07-17	43	52	40,231	—
justageorgiagurl	Georgia	2026-08-11	22	8	6,530	—
karonde.yaps	Georgia	2026-06-24	112	92	210,091	yes
lijaingeorgia	Georgia	2026-08-01	8	17	7,610	yes
livinlifewithjo_	Georgia	2025-11-06	226	13	49,945	—
mads_ventures	Georgia	2019-10-11	11,100	100	6,014,387	—
mayor.tooliani	Georgia	2026-03-10	58	0	0	—
news.with.whit	Georgia	2026-08-10	9	8	4,150	yes
newsfeedfemi	Georgia	2026-07-04	95	104	184,842	yes
paigerants3	Georgia	2026-07-27	383	53	208,849	yes
peachy.topics	Georgia	2026-07-24	36	68	38,012	yes
punx.news	Georgia	2026-07-21	38	22	16,138	yes
rantswithbec	Georgia	2026-08-12	0	0	0	—
ryansays12	Georgia	2026-07-10	29	45	26,734	yes
tashaingeorgia	Georgia	2026-07-31	50	42	48,845	yes
teanews5	Georgia	2026-07-18	60	33	191,356	—
the_shep_report	Georgia	2026-06-03	144	96	60,202	yes
toolieganews	Georgia	2026-06-03	556	223	783,026	yes
treysgeorgiathoughts	Georgia	2026-07-09	132	78	542,149	yes
warmmorningtalks	Georgia	2025-05-02	40	95	36,959	yes
yap4georgia	Georgia	2026-07-23	70	30	24,246	—
itsjenntucker	Iowa	2026-05-13	273	78	320,909	—
just.a.maine.mama	Maine	2026-07-31	39	24	37,607	yes
bergen.ellis	Michigan	2026-05-22	56	37	21,006	—
jennopalis	Michigan	2025-09-05	677	221	3,935,096	—
lily.cather	Michigan	2026-05-22	104	30	50,255	—
rants.with.eliag	Michigan	2025-11-23	131	38	62,143	yes
rants.with.s	Michigan	2026-05-24	19	38	19,486	yes
redhead.rachel	Michigan	2026-02-12	15	0	0	—
annacole5950	N. Carolina	2026-02-23	1,456	257	1,460,144	yes
emma.smith.733	N. Carolina	2026-05-18	1,228	295	1,387,495	yes
hunterrrr972	N. Carolina	2026-01-03	333	330	777,471	yes
junetalks24	N. Carolina	2026-05-21	107	42	123,534	yes
katie.saves20	N. Carolina	2026-02-27	149	60	207,647	—
ridgesidehustle	N. Carolina	2025-08-19	198	8	54,165	—
tyrants51	N. Carolina	2026-05-18	22	39	37,385	—
derektalkss	New York	2026-07-02	5	6	4,704	—
its.emma.in.ny	New York	2026-07-30	23	44	25,679	yes
rae.luvs.ny	New York	2026-05-22	18	67	56,429	yes
rants.with.ki	New York	2026-07-29	4	14	7,351	yes
rants.with.p	New York	2026-07-31	5	28	15,187	yes
rants.withki	New York	2026-08-04	6	22	17,379	—
austin.ohiofied	Ohio	2026-07-21	19	16	32,981	yes
charlie.crestwell	Ohio	2026-08-05	68	27	29,701	yes
chels.chats	Ohio	2026-08-12	1	0	0	—
discount.cole	Ohio	2026-06-30	72	31	65,097	—
evan.williamss55	Ohio	2026-08-05	9	11	7,129	yes
graceisgabbing	Ohio	2026-06-17	208	65	310,594	—
jazzmyneohio	Ohio	2026-07-02	92	76	119,901	yes
kelsie.explains	Ohio	2021-05-25	34	17	9,332	yes
lexigray181	Ohio	2026-07-28	109	18	47,291	yes
mia.laurent16	Ohio	2026-07-31	43	30	26,367	yes
miechel.talks	Ohio	2026-07-30	37	30	27,892	yes
mintinohio	Ohio	2026-07-14	85	24	22,282	yes
newswithbri	Ohio	2026-07-02	86	0	0	—
newswithchloe	Ohio	2026-08-02	45	29	61,829	yes
newswithjob	Ohio	2026-06-24	54	99	89,388	yes
nnicoletalks	Ohio	2025-10-26	920	98	517,752	—
ohio.by.myka	Ohio	2026-07-15	74	94	63,553	yes
ohioinfo	Ohio	2026-07-10	1,848	39	590,907	—
oliviainohio	Ohio	2026-07-06	488	137	814,962	yes
rants.with.kay	Ohio	2026-07-30	91	20	40,268	yes
rants.with.rae	Ohio	2026-06-22	39	24	36,691	—
rantswit.breezy	Ohio	2026-08-12	0	6	2,904	yes
rantswithivy	Ohio	2026-08-03	82	22	107,259	yes
renee.ohio	Ohio	2026-02-19	22	4	5,267	yes
sophia.ranting	Ohio	2026-07-30	234	34	251,103	yes
talia.smith333	Ohio	2026-08-01	90	30	22,115	yes
whocanaffordthis	Ohio	2026-08-03	42	19	16,302	—
addy.sam24	Pennsylvania	2026-02-19	222	157	517,199	—
kailyn.moyer	Pennsylvania	2026-05-14	1,765	312	1,331,331	yes
macyranting	Pennsylvania	2026-06-16	54	215	266,522	yes
rants.with.aimee	Pennsylvania	2026-02-20	533	97	508,824	yes
amira.yaps6	Texas	2026-05-01	229	0	0	—
bellacruz6339	Texas	2026-05-20	201	71	431,280	—
biancalikestoyap	Texas	2026-05-15	50	38	116,855	yes
dayna.bh	Texas	2026-05-14	97	77	291,464	yes
faysbigworld	Texas	2026-07-28	17	55	20,426	yes
harvtalks	Texas	2026-05-14	3	7	4,779	—
heyitssarahtx	Texas	2026-05-20	95	67	147,581	—
jill33112	Texas	2026-04-29	304	135	349,843	—
rants.withgloria	Texas	2026-04-28	22	42	297,219	yes
talkingwfabs	Texas	2026-05-17	20	0	0	—
talkingwithh	Texas	2025-07-25	1,376	350	9,996,653	yes
talkswithharv	Texas	2026-05-22	56	67	284,448	yes
texasaintcheap	Texas	2025-12-13	79	128	174,827	yes
texasgrlspam	Texas	2026-04-13	2,134	90	5,239,059	yes
this.is.taylor2	Texas	2026-05-17	118	67	204,648	—
tiffany.lakes	Texas	2026-05-21	86	46	724,556	—
eliserants44	Wisconsin	2026-07-19	40	41	35,230	—
erin.d.930	Wisconsin	2026-02-19	1,124	336	1,664,253	yes
lifewithjacob1	Wisconsin	2026-05-16	211	276	490,399	yes
rants.with.k89	Wisconsin	2026-05-15	410	112	334,359	—
rantswithsaum	Wisconsin	2022-11-13	1,206	50	651,594	—