Research methodology

Evidence for decisions at signal speed.

How TrendsAGI models event velocity, calibrates thresholds, and routes structured alerts—so teams shorten time-to-action without giving up verification quality.

Start free See operations workflow

Signal definitions

Make every trigger explainable.

Cultural-moment detection is treated as a streaming decision system. Each metric is explicit enough to audit, tune, and map to a deterministic action.

MetricFormalizationPurposeImplementation note
Velocitymentions/hour or events/minuteMeasures current spread intensityPrimary trigger axis for monitor/notify/act tiers
Accelerationd(velocity)/dt over rolling windowsSeparates stable high-volume chatter from breakout movementEscalate if acceleration remains positive across two windows
Source Diversityentropy over source clustersReduces single-source amplification noiseRequire minimum distinct cluster count before high severity
Persistenceminutes above thresholdImproves precision by rejecting brief spikesGate action tier until sustained for N minutes
Anomaly Scoredeviation vs adaptive baselineDetects out-of-pattern behavior per topicRoute to triage queue if anomaly score exceeds configured bound

Detection pipeline

Six stages from observation to learning.

A structured pipeline keeps detection explainable and separates signal computation from downstream automation.

  1. Stage 01

    Observe

    01

    Collect source events and metadata from monitored channels.

    Next · Normalize

  2. Stage 02

    Normalize

    02

    Apply dedupe, token mapping, timezone alignment, and entity extraction.

    Next · Detect

  3. Stage 03

    Detect

    03

    Compute velocity, acceleration, persistence, and anomaly deltas.

    Next · Route

  4. Stage 04

    Route

    04

    Apply tenant rules for severity, category, and destination endpoint.

    Next · Verify

  5. Stage 05

    Verify

    05

    Validate signature + freshness before automation consumes the payload.

    Next · Learn

  6. Stage 06

    Learn

    06

    Review false positives, update thresholds, and adapt to concept drift.

Evidence matrix

Connect literature to product decisions.

Cross-domain research is used as methodological support and explicitly caveated when applied to marketing workflows.

FindingSourceProduct implicationImplementation note
Collective attention accelerates over time; response windows compress.[R1]Detection latency must be minimized for campaign relevance.Favor low-latency ingestion and early-stage triggering.
Real-time message timing improves engagement outcomes.[R2]Operational timing is a product capability, not only a content problem.Optimize time-to-brief and time-to-first-publish SLAs.
Newsjacking effectiveness depends on topical fit and speed.[R3]Alerting must include context, not only counts.Attach rationale fields in webhook payloads for triage confidence.
Social streams can act as real-time event sensors.[R4]Event-detection methods transfer to marketing moment sensing.Treat transfer as methodological support, not direct causality proof.
Burst detection in streams identifies emerging attention states.[R5]Burst models can power initial detect and escalation boundaries.Use burst + persistence together to improve precision.
Concept drift requires continuous threshold adaptation.[R6]Static thresholds decay as topic baselines move.Run periodic calibration from accepted/rejected alert outcomes.
Incident response guidance emphasizes playbooks and verification.[R7]Webhook alerts should map to explicit response runbooks.Route by severity to deterministic workflows.

Transfer caveat: references [R4]-[R7] originate from event detection, stream mining, and incident response. They support detection architecture and response design—not direct marketing outcome equivalence.

Threshold calibration

Make the precision tradeoff explicit.

Start conservatively, then tighten monitor, notify, and act boundaries using accepted and rejected alert outcomes.

TierSuggested conditionObjectivePrecision/recall tradeoffAutomated action
Monitorvelocity >= 3x baseline OR anomaly >= mediumHigh recall, lower precisionCaptures weak signals; accepts analyst review loadLog signal and append context to monitoring queue
Notifyvelocity >= 10x baseline AND persistence >= 15mBalanced precision/recallFilters short-lived bursts while preserving early detectionNotify channel owner and create brief draft
Actvelocity >= 25x baseline AND diversity >= thresholdHigher precision, lower false escalationMay miss very early single-cluster eventsTrigger webhook-runbook automation with assignee routing

Webhook contract

Verify before automation acts.

Payloads are structured for deterministic routing. Consumers verify signature, timestamp freshness, and schema compatibility before taking action.

01

Verify

Validate the HMAC signature and reject stale timestamps.

02

Route

Map tier and category to a queue, owner, and runbook.

03

Act

Create the task and capture acknowledgement latency.

Signed webhook · JSON
POST https://your-webhook-endpoint
X-Webhook-Signature: <hmac-sha256>
X-Webhook-Timestamp: <unix-epoch-seconds>
Content-Type: application/json

{
  "trigger_type": "trend_velocity",
  "tier": "notify",
  "trend_name": "airport closure + storm",
  "velocity": 2400,
  "acceleration": 1.8,
  "source_diversity": 0.74,
  "persistence_minutes": 19,
  "sentiment": "neutral",
  "timestamp": "2026-02-05T18:42:00Z",
  "summary": "Closure chatter accelerated across Northeast travel sources.",
  "source_link": "https://trendsagi.com/blog"
}

Citations

Trace every design choice to its source.

Persistent DOI links keep implementation reviews and threshold-design decisions grounded in the underlying research.

  1. [R1]

    Lorenz-Spreen et al. (2019), Accelerating dynamics of collective attention.

    https://doi.org/10.1038/s41467-019-09311-w
  2. [R2]

    Chae et al. (2025), Real-time marketing messages and consumer engagement.

    https://doi.org/10.1016/j.jbusres.2025.115266
  3. [R3]

    Angell et al. (2019), Newsjacking-based content effectiveness.

    https://doi.org/10.1108/ITP-04-2019-0177
  4. [R4]

    Sakaki et al. (2010), Earthquake shakes Twitter users.

    https://doi.org/10.1145/1772690.1772777
  5. [R5]

    Kleinberg (2002), Bursty and hierarchical structure in streams.

    https://doi.org/10.1145/775047.775061
  6. [R6]

    Gama et al. (2014), Survey on concept drift adaptation.

    https://doi.org/10.1145/2523813
  7. [R7]

    NIST SP 800-61r3 (2025), Incident response recommendations.

    https://doi.org/10.6028/NIST.SP.800-61r3

Operationalize the detection model.

Start with one watchlist, one endpoint, and explicit monitor, notify, and act tiers. Refine thresholds from observed precision and recall.

Start free View API docs