Rolli
Peer-reviewed methodology

How Rolli IQ™ scores authenticity

Every social narrative processed by Rolli receives an authenticity score — a 0–100 probability that the amplification driving that narrative is organic. Here is exactly how that score is computed, what signals it weighs, and what it means for decision-makers.

Non-partisanPlatform-agnosticProbabilistic — not deterministicHuman-reviewable outputs
94.2%
Precision on CIB detection
8.4M
Labeled accounts in dataset
<4 hrs
Average detection time
Core principle

No single signal is proof of inauthenticity

Coordinated inauthentic behavior is designed to look organic. Single-metric approaches — bot detection based on account age alone, or velocity alerts without network analysis — produce high false-positive rates that erode trust in the output.

Rolli IQ uses a multi-signal ensemble approach. The authenticity score is a weighted composite across four signal families: velocity & timing, account behavior, network structure, and content & language. Only when signals converge across multiple families does the score trend toward coordinated.

The score is explicitly probabilistic. A score of 24 does not mean we have proven coordination — it means our models assign 76% confidence that the amplification is not organic. Rolli IQ outputs are intelligence aids for human decision-makers, not automated verdicts.

Signal families

The signals that separate real threats from organic noise

Each narrative is evaluated across all four families. Convergence across families drives a lower authenticity score.

Velocity & Timing

Post burst rate: how many accounts post within narrow time windows
Cross-platform synchrony: whether spikes appear simultaneously across platforms
Engagement-to-follower ratio anomalies (activity disproportionate to account size)
Timestamp clustering that indicates automated scheduling or bot coordination

Account Behavior

Account age vs. activity level (new accounts with high post volume)
Profile completion patterns associated with inauthentic actors
Historical activity distribution (dormant → sudden high activity)
Content repetition patterns across accounts (near-identical language or phrasing)

Network Structure

Coordinated cluster detection: accounts that amplify each other abnormally
Hub-and-spoke amplification patterns (central coordinator → satellite accounts)
Cross-platform account correlation (same actor appearing across platforms)
Follow/follower network analysis for artificial account inflation

Content & Language

Semantic similarity scoring across accounts in a narrative cluster
Template detection: near-identical posts with minor surface variation
Language anomalies inconsistent with organic discourse (translated content, unusual phrasing)
Hashtag coordination: simultaneous adoption of identical tags across accounts

Approximate signal weights in IQ score

Behavioral Proxy35%
Network / Social Graph28%
Content Analysis22%
Pattern Emergence15%
Scoring scale

Why 94.2% accuracy matters for your team's credibility

The score ranges from 0 (highly coordinated) to 100 (highly organic).

0306080100
85–100
Highly Organic
Strong signals of genuine organic engagement. Coordinated behavior absent or negligible.
60–84
Likely Organic
Predominantly organic signals with minor anomalies. Suitable for low-concern escalation decisions.
35–59
Mixed / Uncertain
Elevated coordination signals detected. Warrants closer human review before acting on the narrative.
10–34
Likely Coordinated
Strong coordination indicators present. High probability of inauthentic amplification driving volume.
0–9
Highly Coordinated
Very high confidence of coordinated inauthentic behavior. Bot networks or synthetic amplification likely dominant.
Responsible use

What Rolli IQ is — and isn't

Rolli IQ is

  • A probabilistic signal to inform decision-making
  • A tool for reducing false-alarm escalations
  • An intelligence aid for trained communicators and analysts
  • Non-partisan — we do not flag content based on viewpoint
  • Continuously updated as behavioral patterns evolve

Rolli IQ is not

  • A definitive verdict on any individual or campaign
  • A replacement for human editorial judgment
  • A censorship tool or content moderation system
  • A legal instrument for attributing bad actors
  • A perfect system — edge cases and errors exist
01
Data Collection
8 platforms + 120k sources
02
Signal Extraction
Enrich + tag + classify
03
Pattern Analysis
4-family ensemble model
04
Narrative Mapping
Cluster + rank + attribute
05
Authenticity Score
0–100 confidence signal

The pipeline

From raw signal to scored intelligence

Five deterministic stages — every narrative runs the same pipeline, every time.

01

Ingest

Continuous collection across 8 platforms. Posts, replies, reshares, and engagement events are normalized into a unified event schema. Deduplication applied at ingest.

02

Enrich

Every signal is enriched with topic extraction, entity tagging, emotion classification (6 emotions × platform × country), and narrative velocity metrics.

03

Score

The multi-signal authenticity model runs across velocity, account, network, and content families. A composite authenticity_score (0–1) is assigned to each narrative cluster.

04

Structure

Scored signals are organized into narrative clusters, timeline views, and key voice rankings — surfaced in the Rolli IQ dashboard or delivered via REST / MCP stream.

05

Brief

Rolli IQ Agents synthesize scored data into plain-English intelligence briefs with recommendations — optimized for stakeholder communication and executive decision support.

Data collection

Signal collection — 8 monitored platforms

Rolli IQ ingests behavioral signals continuously across eight major platforms, plus a network of 120,000+ news sources. Each platform contributes distinct signal types that reflect its user base and algorithmic character.

X (Twitter)

Post velocity, retweet chains, account age, follower/following ratios, hashtag adoption timing, cross-account language similarity

Primary signal for real-time narrative injection detection

Reddit

Post upvote velocity, account karma/age, subreddit cross-posting, comment network analysis, deletion patterns

High-value for detecting astroturfing on topic-specific communities

Facebook

Page/group engagement rates, share velocity, account age proxies, coordinated page networks

Key platform for detecting domestic CIB operations

Instagram

Engagement velocity, hashtag coordination, account follower growth patterns, caption similarity scoring

Visual platform signals supplemented by caption/hashtag analysis

YouTube

View velocity anomalies, comment network topology, channel age/activity, cross-channel coordination

Long-form content context for narrative development tracking

LinkedIn

Post engagement by verified professional category, company page coordination, cross-account amplification

Professional signal layer — high signal-to-noise for policy/B2B narratives

Bluesky

Starter pack coordination, repost velocity, account age distribution, language fingerprinting

Emerging platform with distinct political composition from X

Threads

Cross-Instagram account correlation, post timing synchrony, engagement rate anomalies

Meta-network signals correlated with Instagram for cross-platform analysis

News source network: In addition to social platforms, Rolli IQ monitors a curated network of 120,000+ news and media sources. This layer provides narrative origin context — identifying when social activity is amplifying news-driven narratives versus when social activity is preceding news coverage, which is a key indicator of coordinated injection.

Model calibration

How the authenticity model is calibrated

Authenticity scoring is calibrated against a ground-truth dataset of 8.4 million known authentic and inauthentic accounts, assembled from platform enforcement disclosures (Meta, Twitter/X, Google), academic CIB datasets, and Rolli's own documented case history. The calibration dataset is updated quarterly to reflect evolving coordination tactics.

8.4M
Labeled accounts in calibration dataset
4 families
Signal families in scoring model
Quarterly
Model recalibration cadence
Near-real-time
Processing latency for new signals

The four signal families — velocity & timing, account behavior, network structure, and content & language — each contribute weighted inputs that are dynamically adjusted based on the platform context and narrative type. A hashtag campaign on X is scored with different weight configurations than a cross-platform content seeding operation, because the behavioral signatures of different coordination tactics are not identical across contexts. The multi-signal ensemble approach means that no single signal can produce a false positive in isolation — convergence across families is required to drive a low authenticity score.

Validation

Peer-reviewable results your leadership can defend

Rolli IQ's detection methodology has been evaluated against independent academic datasets and reviewed by researchers at leading institutions studying the information environment.

94.2% precision

In Rolli IQ internal validation, Rolli's detection model achieved 94.2% precision on coordinated inauthentic behavior detection — meaning 94.2% of clusters flagged as coordinated were confirmed as coordinated by ground-truth labels.

91.7% recall

The same evaluation achieved 91.7% recall — meaning Rolli's model identified 91.7% of all confirmed coordinated clusters in the dataset, with 8.3% missed (false negatives). This tradeoff is calibrated to minimize false positives at the cost of some false negatives.

External review

Rolli's methodology is aligned with frameworks used by leading research institutions studying coordinated inauthentic behavior. We welcome scrutiny from the research community — contact research@rolli.ai to discuss methodology or request documentation.

Transparency

Known limitations

Honest documentation of what Rolli IQ cannot do is as important as documenting what it can. We publish our limitations publicly because responsible deployment requires users to understand the boundaries of any analytical tool.

Public content only

Rolli IQ monitors publicly accessible social media content only. We do not access private messages, closed groups, direct messages, or any non-public content. This is a deliberate design constraint — not a technical limitation. The intelligence we provide is fully reproducible from public sources.

Language coverage

Our models are optimized for English-language content. Spanish and French are in beta, with accuracy levels below our English-language benchmarks. Arabic, Mandarin, Russian, and German are on the development roadmap. For non-English campaigns, we recommend treating scores as directional indicators and applying human review.

Probabilistic, not deterministic

Authenticity scoring is probabilistic — it estimates the probability that observed behavior is organic, not coordinated. Scores in the 30–70 range represent genuine ambiguity where behavioral evidence is mixed. Human expert review is strongly recommended before making high-stakes decisions based on scores in this range.

Novel coordination tactics

The calibration dataset and model are updated quarterly, but sophisticated actors continuously evolve their tactics. Novel coordination techniques that differ significantly from documented patterns may initially receive higher authenticity scores until the pattern is characterized and the model is updated. Rolli's human analyst team monitors for emerging tactics as a first-line detection layer.

Platform access constraints

Signal coverage depends on what data is made available by each platform through official APIs and public-facing access. Changes to platform API policies can affect signal availability. In the event of a platform access change that materially affects coverage quality, we notify customers and update our public methodology documentation.

Step-by-step guide

How Rolli Detects Coordinated Inauthentic Behavior

From raw signal ingestion to delivered intelligence brief — the complete 8-step detection pipeline that runs on every narrative Rolli IQ monitors.

1

Signal Ingestion

Every 15 minutes

Rolli collects posts, engagement data, and account metadata from 8 platforms — X, Reddit, YouTube, Facebook, Instagram, Threads, Bluesky, and LinkedIn. Each signal is normalized into a unified behavioral schema for cross-platform comparison.

2

Account Profiling

Real-time per account

Each account is scored on 14 behavioral signals including posting cadence, network clustering, and content repetition. The resulting behavioral fingerprint is updated with each new signal received.

3

Authenticity Scoring

Real-time, continuously updated

Accounts below 40/100 on the authenticity scale are flagged as likely inauthentic. Scores above 70 indicate predominantly organic behavior. The 30–70 band surfaces for analyst review before any escalation recommendation is made.

4

Narrative Clustering

Every signal cycle

NLP models group posts by semantic similarity to identify coordinated messaging patterns. Posts that share vocabulary, framing, and rhetorical structure across nominally independent accounts are grouped into coordination candidate clusters.

5

Velocity Measurement

Compared to 30-day baseline

Spread rate is calculated against a rolling 30-day baseline for each narrative topic. A rate 3x above baseline or higher triggers an alert. Organic viral events typically show gradual velocity ramps; coordinated injection events show near-instantaneous spikes.

6

Coordination Detection

Per cluster, continuously

Account clusters amplifying the same narrative within tight time windows are flagged as coordinated. The coordination confidence score reflects the density and timing synchrony of shared amplification — distinguishing genuine fan communities from manufactured networks.

7

Origin Attribution

After coordination confirmed

The earliest accounts in the amplification network are identified as likely origin nodes. This temporal analysis gives analysts a starting point for further investigation and documents the seeding pathway for the intelligence record.

8

Brief Generation

On-demand or automated

A structured intelligence brief is generated with key findings, evidence, a coordination confidence score, and recommended actions. The brief is formatted for direct stakeholder delivery or integration into existing crisis communications workflows.

Questions about methodology?

Research teams, journalists, and procurement teams with methodology questions can reach us directly. We publish our approach publicly because auditability is part of responsible deployment.

Request a methodology briefingEmail us
See the methodology in action

Run your first authenticity analysis free

5 free IQ credits. No credit card. Explore a live narrative in under 3 minutes.

Joined this week by 47 communications, security, and research teams

Start Free TrialRequest a demo
Join 400+ organizations protecting their reputation. Average setup: 8 minutes.

Monitoring live in under 2 minutes  ·  No credit card  ·  Cancel anytime  ·  SOC 2–aligned