The claim being made is never in the text. It is in the apparent count of people making it.
Why it works
We use consensus as a shortcut, and reasonably so — checking every claim personally is impossible, and "many unconnected people think this" is genuinely useful information when it is true. Sockpuppetry attacks the shortcut rather than the argument.
Against an individual it does something more specific. Being told by one person that you are unreasonable is an opinion. Being told the same thing by nine accounts in an afternoon is an experience that is very hard to hold out against, and the effect does not depend on any of them being persuasive.
Recognising it — and the limits of doing so
Commonly cited signals: accounts created around the same time, thin histories, similar phrasing or rhythm, synchronised arrival, no social graph, interest that does not extend beyond the one dispute.
Two honest caveats.
None of these is proof, and several are ordinary. New accounts exist. Quiet accounts exist. People who read the same sources sound alike. Anyone who has been wrongly accused of running alt accounts will tell you how cheap the accusation is and how impossible it is to disprove.
The traditional signals are weakening. Older detection advice leaned on things like poor grammar and repetitive phrasing. Text produced with a language model has neither, so content-level tells are worth much less than they were. What is left is behaviour and coordination — timing, overlap, what accounts do rather than how they write — and that is genuinely harder to see from inside a conversation.
What to do
Do not accuse. It cannot be proven from where you stand, it is the single most effective way to make yourself look paranoid to onlookers, and if you are wrong you have done to someone else exactly what was being done to you.
Report the pattern to the platform, which can see account metadata you cannot. Then disengage from the volume — not because silence resolves it, but because replying to nine accounts is a use of your evening with no available win.