01
Set up the task
Five groups of settings — direction, engines, materials, review & publishing, collaboration — all set by a human; the two placeholder switches are honestly labeled as not yet changing generation behavior.
Executed by: human
Insights Content quality control
From idea to publish, 12 gates stand in between
A 6-stage pipeline, 12 review gates and a 10-node autopilot route — AI content production is where trust is hardest to earn, so we lay the whole path open.
This page covers how many gates exist, where they sit and who rules on them; the judging rules inside each gate (thresholds, word lists, model prompts) are not published.
Six-stage pipeline
From task setup to verified results, nothing skipped
Every stage names its executor — human / AI / mixed / system — so you know exactly who owns each step.
02
AI generation
Schedule → pick title → assemble prompt → retrieve knowledge → write → illustrate → summarize → tone & public-order check → GEO score → store.
Executed by: AI
03
Review
Content passes risk gating, brand guideline, public-order review, industry compliance pack and human review in turn; anything failing reverts to draft — never silently published.
Executed by: human + AI
04
Publish
Three routes (direct API / manual list / desktop publishing), six states, and triple-layered retry on failure.
Executed by: system
05
AI visibility monitoring
Real multi-engine sampling, fixed question sets re-tested, competitor comparison and source analysis — trends built weekly.
Executed by: AI
06
Verify results
Measurement-condition snapshots plus a "false improvement" self-check: numbers can be recomputed and every change can be explained.
Executed by: human + AI
Review gates
12 gates, in their real execution order
The table below follows the actual order in code — not a marketing checklist. Every gate maps to a real implementation, cross-checked against the codebase.
| Gate | Ruled by | Applies at | Power |
|---|---|---|---|
| 01 Keyword cannibalisation check | Rule engine | At generation | Blocks (reverts to draft) |
| 02 Sensitive-word & risk scan | Rule engine | At generation | Blocks (reverts to draft) |
| 03 Risk gate | Rule engine | At generation | Blocks (reverts to draft) |
| 04 Brand guideline check | AI review | At generation | Blocks (reverts to draft) |
| 05 Public-order & morality review | AI review | At generation | Blocks (reverts to draft) |
| 06 Custom skill gate | AI review | At generation | Blocks (reverts to draft) |
| 07 Platform compliance pre-check | Rule engine | At publishing | Blocks (reverts to draft) |
| 08 Industry compliance pack | Rule engine | At publishing | Blocks (reverts to draft) |
| 09 Human review | Human | At generation | Blocks (reverts to draft) |
| 10 Client confirmation | Human | At generation | Optional (client sign-off) |
| 11 Autopilot pre-check | Rule engine | Every path | Blocks (reverts to draft) |
| 12 AI content label | System | Every path | Forced (cannot be bypassed) |
The "AI content label" gate is mandatory: the system injects it on every publishing path and it cannot be bypassed — meeting China's labeling rules for AI-generated content.
Autopilot route
10 unattended nodes — and where a human must approve
Turning on autopilot is not letting go: key nodes on the route are human checkpoints, and nothing moves until they sign off.
AI visibility sampling
Asks each engine, for real, whether your brand comes up — producing the visibility rate and the gap keywords you should have appeared for.
Effect audit
Gathers the three checks run on every sample (factual accuracy, citation fidelity, visibility rate) so "redo it if the effect falls short" can be expressed in the route.
Keyword mining
Expands the gap keywords into a batch of question phrases that can be written directly into articles.
Content generation
Creates and runs tasks automatically, through exactly the same generation chain as a manually created task.
Compliance check
Scans generated content for compliance; hitting a red line sends the whole run to human review instead of auto-releasing it.
GEO scoring
Scores each article in the batch for GEO readiness, so later steps can route or prioritize by score.
Content review Human checkpoint
Human checkpoint: in full-autopilot mode a deterministic pre-check runs first and only an all-clear auto-releases it; anything else goes to a person.
Distribution
Queues approved content for distribution to the channels and accounts you have connected.
Read-back
Reads platform-side metrics back for published content, so performance lands back in the platform.
Finish
Summarises the run's output, closes it out and keeps a complete execution record.
Every step leaves an audit trail; the review node is a human checkpoint (nothing moves until confirmed). Full-autopilot and semi-autopilot modes, chosen by how much your team wants to stay in the loop.