Build an Evidence-Backed Content Agent

Reliable content automation separates research from writing and publishing. Agent should prepare evidence-backed changes, then open a pull request for human review.

This article explains architecture and editorial reasoning. Use executable content-agent workflow to run process, or index record for compact machine-readable definition.

Google permits AI-assisted content, but warns that generating many pages without added user value can violate scaled-content policy. Production system therefore needs rejection rules, not only writing prompts.

Pipeline#

Opportunity
→ approved brief
→ evidence map
→ draft
→ claim verification
→ editorial review
→ SEO/AEO review
→ pull request
→ human approval

Each stage produces an artifact another stage can inspect.

1. Opportunity#

Opportunity record should answer:

• Which reader problem exists?
• Which query or task expresses it?
• What current pages fail to provide?
• What original asset can this page add?
• Which existing page might it duplicate?

Reject opportunities based only on keyword variation. "Best AI email tools for startups" and "best AI email tools for small companies" may describe same decision.

2. Brief#

Approved brief defines scope before research expands:

primary_intent: choose architecture for repeatable AI marketing work
reader: marketer building first repository-based workflow
required_asset: runnable folder structure and approval model
must_answer:
  - when to use prompt, skill, workflow, or agent
  - where evidence lives
  - who can publish
exclusions:
  - unsupported performance claims
  - vendor rankings

3. Evidence map#

Research agent records claims before prose:

Source type matters. Vendor documentation supports what vendor says product does. It does not equal independent performance testing.

4. Draft#

Writing agent receives approved brief and evidence map. It should not browse new sources during drafting because hidden research makes verification harder.

Draft requirements:

• Direct answer near top
• One distinct reader intent
• Original template, example, test, or method
• Source links near claims
• Clear uncertainty
• No invented quotes, numbers, or product behavior

5. Verification#

Verification agent checks every factual sentence against evidence map.

Possible outcomes:

• Supported
• Vendor-documented
• Inferred
• Stale
• Contradicted
• Unsupported

Unsupported claim gets removed or returned to research. Writer cannot quietly soften it into vague wording.

6. SEO and answer review#

Review technical basics:

• Stable canonical URL
• Useful title and description
• Clear headings
• Crawlable links
• Internal topic relationships
• Structured data matching visible page
• Accurate publication and update dates
• Short direct answer for question-led pages

No special markup guarantees inclusion in an AI answer. Strong source clarity and useful page structure improve machine retrieval without pretending to control third-party systems.

7. Pull request#

Agent opens one focused pull request containing:

• Article
• Evidence map
• Related-link updates
• Generated index changes
• Validation report

Reviewer sees what changed and why. Agent cannot merge its own work.

Publish gate#

Publish only when page passes five tests:

1. Distinct intent
2. Verified factual claims
3. Original value beyond source summaries
4. Relevant internal relationships
5. Named human approval

Production volume should follow passing pages, not precede them.