Real interest scatters. A push fires together.

People who care about a topic post at their own pace. Accounts acting together do the same thing within seconds of each other, and do it again.

For each trending topic, every five minutes, the product looks for accounts that did the same thing within a short window, keeps the pairs that did so more often than chance would explain, and finds the groups they form.

The result is a coordination report for the topic: the groups, their accounts, the evidence for each, a score and one of four labels: organic, unclear, lead or likely coordinated.

Media Verdict console

Six ways of acting together

Each time two accounts in a topic do the same thing inside the window, the pair is counted. Something that very many accounts are sharing counts for less, so one viral link or hashtag cannot make a network on its own. The windows below are the default settings.

  1. 01

    The same repost

    Two accounts repost the same post within 60 seconds, by default.

  2. 02

    The same text

    The same words, or a rewording close enough in meaning, in the same language or across languages, within five minutes by default. Captions reused on short videos count as text.

  3. 03

    The same link

    The same link within 60 seconds, by default.

  4. 04

    The same hashtag

    The same hashtag within 30 seconds, by default.

  5. 05

    The same hashtags in the same order

    A sequence of three or more hashtags in the same order within five minutes by default. It is the trace that holds on video platforms, where there is little text to compare.

  6. 06

    The same media

    The same image, video or audio, matched by its fingerprint, within five minutes by default.

From pairs to a named group

Pairs become groups in three steps, and each step is written into the report.

In the drawing, each square is an account in the topic. Grey accounts belong to no group; black ones are members, joined by a short line where two of them acted together. Groups 1, 2 and 3 are bracketed beneath, and group 2 is marked in red as the one being examined.

Media Verdict console
1 2 3
Illustration
  1. 01

    Pairs

    Two accounts that did the same thing within the window, counted for each of the six ways.

  2. 02

    Chance removed

    A pair that met once is dropped. A pair is kept only when it acted together well beyond what a statistical model of ordinary activity predicts. Accounts left with no pair drop out.

  3. 03

    Groups

    Communities are found within each of the six ways and across all of them together. A group needs at least five accounts and at least one in twenty of the topic’s posts. A text group with five or more different rewordings is marked as a campaign of rewordings, with its languages.

  4. 04

    No group found

    If no community qualifies, the whole topic is scored as one group, by the same method.

Who is pushing this

The Accounts screen ranks the accounts behind a topic, a body you track, a group or a time window.

Media Verdict console
1 2 3 4
IllustrationNumbers match the four parts beneath. Each row is one account, ranked by the posts it put into the topic; names, counts and the rising and repeat marks are left blank, except one repeat account in red.
  1. 01

    Ranked by reach or by posts

    For each account: platform, group, posts and reach in the window, topics, groups joined and when it was first seen.

  2. 02

    Rising accounts

    Accounts whose reach is climbing against their own seven-day median.

  3. 03

    Repeat accounts

    Accounts found in groups in two or more earlier topics within 90 days, across your cases.

  4. 04

    Kept inside

    Account-level rows stay in internal reports. They never appear in an external one.

The organic share

0 1
Illustration

Each account that posts in a watched topic gets an authenticity score from its behaviour alone: how often and how regularly it posts, when it was created and whether it went quiet and came back, whether follows and replies go both ways, whether its picture matches known stock or generated images, and the pattern of its handle.

The organic share is the share of the topic’s posts with each post weighted by that score. It is shown on each topic’s lane in Timeline, and stance shares are weighted by it.

It is a weight, never a filter. Nothing is removed, the raw counts stay beside the weighted ones, and the score never decides a label. Until it has been calibrated it is marked as not validated.

What a finding contains

PartWhat it holds
AccountsThe members of each group, with checks on each member for the reviewer: an account created inside the group’s creation window, a profile picture found elsewhere by reverse image search, posting hours against stated location, posts outside the campaign, and the share of replies that stay inside the group.
EvidenceThe pairs that tie each group together, by way of acting together, with sample posts; and eleven measures from 0 to 1: synchronised bursts, text templates, aligned narrative, repeat offenders, cross-platform spread, co-repost density, schedule regularity, creation cluster, new or small accounts, repost-only accounts and inter-arrival gaps. A measure whose inputs are missing is left empty and the report says so.
Label and scoreOrganic, unclear, lead or likely coordinated, with the score. A score at or above the upper line reads likely coordinated only when at least three of the eleven measures pass their own mark; with fewer it is a lead. Between the two lines, three or more make a lead and fewer leave it unclear; at or below the lower line it is organic.
Time windowThe topic’s window and, for each way of acting together, the seconds inside which two actions count as together. Windows from ten seconds to fifteen minutes are tried in the same run, and the result is recorded.
How to reproduce itThe method is stored with the finding: the ways of acting together used, the windows, the statistical cut-off, the thresholds and the configuration version. The members and pairs can be exported, and a run from them gives the same report.

What it does not prove

A finding shows that accounts acted together: the same action, within seconds, far more often than chance. It does not show who is behind the accounts.

A high score alone is not a finding. A topic whose score is at or above the line for likely coordinated is called so only when at least three of the eleven measures pass their own mark; with fewer it is shown as a lead. A score between the two lines is a lead with three or more, and unclear with fewer.

The account score is a weight and an aid for the reviewer. It never decides a label and never appears in an external report at account level.

Coordination is read from public posts. A campaign organised on WhatsApp is seen only when it reaches public platforms. The answers that AI assistants give are kept out of coordination entirely.

The score becomes a probability once it has been calibrated on at least 50 topics labelled by two of your analysts whose labels agree closely enough. Until then it is shown as a score.

Questions

Does it say who is behind a campaign?

No. It shows which accounts acted together, how and when, with the evidence. Who directs them is a question for your analysts, and a finding does not answer it.

Can a popular post look coordinated?

A link or hashtag that very many accounts share counts for less, pairs must act together well beyond chance, and a high score becomes a finding only with at least three of the eleven measures. The account checks and the stored method let a reviewer see why a group was found.

How is a finding checked?

Every group comes with member checks for a reviewer, the method is stored with the report, and a run from the exported members and pairs gives the same report.

How often is it run?

Every five minutes for each trending topic, on the topic as it stood at the previous run.

Bring a push your team suspected was organised, and see what the evidence for it would look like.

Request a demo