Article
Organic or Coordinated: How to Tell a Real Spike From a Manufactured One
A volume chart cannot tell genuine public anger from a seeded push, because both look identical when you only measure how loud the room is. An insider walkthrough of the behavioural signals that actually separate coordinated amplification from organic reaction.
Every serious conversation about narrative monitoring eventually collapses into one question: is this spike real, or did someone build it? Everything downstream, whether you respond, how loudly, and how fast, depends on the answer. And it is the one question a standard monitoring dashboard is structurally unable to answer.
This is the problem our coordination engine exists to solve. This post is an honest walkthrough of how you actually approach it, without the marketing gloss.
Why volume tells you almost nothing
Start with what a normal tool measures. It counts mentions, tracks sentiment, and draws a line that goes up when the conversation gets loud. That line is the thermometer. It is genuinely useful for telling you that something is happening.
It is useless for telling you what. A real groundswell of ten thousand angry customers and a seeded push of five hundred coordinated accounts amplified into ten thousand impressions can draw the exact same curve. Sentiment does not separate them either, because a good manufactured campaign borrows real emotion. If your only instruments are volume and sentiment, the manufactured spike and the genuine one are indistinguishable, and you will respond to both the same way, which is to say wrongly to at least one.
To tell them apart you have to stop asking how loud and start asking how, by whom, and how it is moving. Those answers live in structure and timing, not in volume.
The fingerprints coordination leaves behind
Organic anger and manufactured anger are produced by different processes, and different processes leave different traces. The signal is behavioural.
Genuine reaction is messy. It comes from people who do not know each other, at human speed, in their own words, spread across the day as people wake up, see something, and react. It has no rhythm because no one is conducting it.
Coordinated pushes are conducted, and conductors leave a beat. The signals that matter most in practice are these:
- Synchronized timing. Independent people do not post within the same narrow windows again and again. Coordinated accounts do, because they are triggered together.
- Near duplicate phrasing. Real people describe the same grievance in wildly different words. Manufactured campaigns reuse language, sometimes exactly, sometimes lightly spun, across accounts that claim no connection.
- Sudden cluster formation. Organic movements grow outward through existing social ties. Seeded ones appear as a dense cluster that materializes almost fully formed, which is not how human attention normally spreads.
- Recycled and repurposed media. Old footage relabeled as current, a screenshot with no verifiable origin, a clip from somewhere else entirely. Manufactured narratives lean on reused assets because producing genuine evidence is harder than recycling.
- Amplifier to audience mismatch. A handful of high reach accounts doing the pushing while genuine engagement underneath stays thin is a classic shape. The narrative looks big from the top and hollow at the base.
No single one of these is proof. Real events sometimes produce synchronized posting, and coordinated campaigns sometimes recruit genuine anger. The signal is in the combination and the weight of evidence, which is exactly why this is a modelling problem and not a rule you can write down in an afternoon.
Reading it as a graph, not a feed
The mental shift that makes detection tractable is to stop looking at a feed of posts and start looking at a network of behaviour.
Treat accounts, their timing, their language, and their connections as a graph. Coordinated activity shows up as unusual structure in that graph: tight clusters with synchronized behaviour and shared phrasing that are too regular to be organic. You are scoring the shape of the conversation, not the content of any single post. This is why a heavier account with a long, believable history can still be caught, because coordination is a property of how a group of accounts behaves together, not of any one profile.
Under the hood this means combining several weak signals into one calibrated read, then constantly checking that read against reality so the thresholds do not drift. It is closer to fraud detection than to sentiment analysis, and it should be, because you are detecting manufactured behaviour, not measuring mood.
The two hard parts nobody advertises
Two things make this genuinely difficult, and they are where most tools quietly give up.
The first is language. In India a narrative does not sit politely in one language. It moves through Hindi, English, and code-switched text inside a single thread, and half the phrasing signal lives in exactly the mix that English only models cannot read. If your language coverage is shallow, your coordination detection is shallow, because you cannot see the duplicate phrasing you are supposed to be matching.
The second is closed platforms. The most consequential spread often happens after a narrative jumps from an open platform into private messaging, where you cannot see individual messages and should not try to. The work there is tracking the narrative’s movement toward that layer and reading the public signals that a jump is happening, not surveilling private chats. Getting that boundary right is both an engineering and an ethics problem, and it matters.
From detection to decision
Detecting coordination is necessary but not the point. The point is what you do with it in the next hour.
A read that says this spike is coordinated, driven by these clusters, and accelerating toward an inflection is only valuable if it arrives while there are still hours to act, and if it comes with a recommended move rather than another chart to interpret. That last step, turning a coordination score into a next action and then tracking whether the response worked, is the difference between an analysis tool and a decision tool.
That is the line we drew when building Pulse. Radar is not enough on its own. You want the radar and the order to act.