An autonomous SEO content engine that keeps publishing verified organic content
Azguards built an autonomous pipeline that finds search-worthy topics, researches and drafts them, has a second model audit every claim, and lands each article in the CMS as a draft with an evidence panel. A person always decides what goes live.
148 passing tests
- Pipeline stages
- 6, from plan to schedule
- Writer and auditor
- Two different models
- Human review
- A few minutes per draft
- Publishing
- Human approval required
- Test suite
- 148 passing tests
Most AI blog automation fails in one of two ways
It either produces confident, generic text with invented statistics, or it produces accurate text that never ships, because CMS entry, SEO fields, social posts and approvals stay manual. We set four requirements before writing any workflow.
Four requirements set before writing any workflow
Each requirement shaped a specific design decision in the engine.
Search-led, not guess-led
Every article starts from a researched topic with a target query, not a prompt like write a blog about X.
Verifiable
Every factual claim must trace to a source. Anything the system cannot support is removed, not softened.
End to end
Output lands where the team works: a CMS draft, a review email and a social queue, not a chat window.
Controlled
Non-technical staff can steer, pause and approve without opening the automation engine.
One pipeline, six stages
A weekly SEO planner and a fact-checked drafting pipeline already ran separately. This engine joins them: research feeds verification, and verification feeds review and distribution.
Plan and discover
Weighted content formats and search-grounded topic proposals. The weekly plan is emailed so a human can veto it.
Research and blueprint
A search-capable model builds a sourced fact sheet. Anything unverified is tagged [UNVERIFIED] rather than stated as fact.
Draft and verify
A writer model drafts to a brand voice spec. A different model audits every claim against the research.
Review gate
The article lands in the CMS as a draft with a traffic-light evidence panel. Nothing goes live without a person.
Social rollout
A three-day plan with rendered cards. A number guard blocks any figure, identifier or version not in the article.
Approve and schedule
Reviewers approve each post. The publisher sends only approved rows, at the set local time.
Architecture and design decisions
Six stages share one control plane in Google Sheets. Each decision below exists to keep a person in charge of what gets published.
Decision 01
A model is poor at catching its own invented facts
The writer and the auditor run as two different models, and the engine refuses to run if they are the same model.
Architecture Narrative
The research fact sheet feeds a writer model. A separate auditor model checks each claim: supported claims go to the CMS draft, unsupported claims are removed first.
Decision 02
Automation that can publish is one bad run away from a public mistake
When a claim cannot be verified against the research, the engine removes the sentence rather than softening it. Rewording a wrong claim keeps the error in the text.
Architecture Narrative
The engine creates drafts only. A person reviews each one, then approves it for publishing at the set time or sends it back for edits.
Decision 03
A softened wrong claim is still wrong
The engine creates only drafts and draft rows. Publishing and posting need a person's approval.
Architecture Narrative
Flagged claims are removed from the draft. If a flagged claim cannot be found in the text, the draft is marked red instead of passing through.
What is built, and what is not yet claimed
The engine is running and tested. Some results are still ahead, so we only state what we have measured.
Built and live-tested
The pipeline runs end to end from topic discovery to CMS draft, and the test suite passes.
Not yet claimed
We do not yet report ranking, traffic or revenue results from the engine, so none are stated here.