Tutorial · Automation
Build an evidence-first AI content research pipeline.
Automation is useful when it removes repetitive collection and formatting. It becomes dangerous when it automates the part that needs judgment: whether a product is actually good, whether a claim is supported, and whether a recommendation deserves a reader's trust.
The pipeline
public demand signals → opportunity records → scoring → evidence pack → content hypothesis → draft → factual/claim QA → human approval → publish → clicks/conversions/usefulness signals → learn and reprioritize
This architecture is intentionally boring. The database and evidence pack matter more than the text generator because they preserve what the system knew when a decision was made.
1. Collect demand, not just trends
A spike in mentions is not necessarily buyer intent. Keep separate signal types:
- Search intent: “X vs Y”, “X pricing”, “X alternative”, “is X worth it”.
- Problem intent: repeated questions about a workflow, cost or failure.
- Launch/news: new product, model or feature.
- Community objections: complaints about price, setup, reliability or missing features.
- Commercial signal: affiliate availability, commission structure and target-market access.
The best content opportunities often combine buyer intent with a product you can genuinely test.
2. Store every candidate as a record
A minimal opportunity table might contain:
id discovered_at product_a product_b intent source_urls[] problem_statement target_reader affiliate_available evidence_status score state notes
Use SQLite, Postgres or even a well-structured spreadsheet at the beginning. The important part is persistence: do not let promising ideas disappear inside chat history.
3. Score for revenue and editorial value separately
One score creates perverse incentives. Keep at least two:
- Editorial score: usefulness, evidence availability, originality, ability to demonstrate, audience fit.
- Commercial score: buyer intent, commission economics, conversion likelihood, target-market accessibility and refund/churn risk.
A high commission should never convert a weak product into a positive recommendation. Commercial score determines whether a useful piece can also monetize—not whether it is useful.
4. Build an evidence pack before drafting
The drafting model should not browse randomly and improvise. Give it a compact evidence pack:
- official pricing/terms URL and access date;
- official documentation for material features;
- your direct test notes and screenshots where available;
- known limitations and failed tests;
- recent community objections, clearly labeled as anecdotal;
- alternatives;
- affiliate/disclosure status;
- claims that are explicitly forbidden until verified.
This makes hallucination easier to detect because the permitted factual surface is defined.
5. Generate a hypothesis, not “an article”
A good content hypothesis has a reader, decision and proof:
For: technical solo creators Decision: ElevenLabs or local TTS Proof: same 10 scripts + cost per accepted minute Format: comparison article + 30s short demo Success: qualified outbound clicks + strong completion/saves
Now the content exists to answer a decision, not to fill a publishing calendar.
6. Create a claim ledger
For every material statement, classify it:
| Claim type | Required evidence |
|---|---|
| Current price/feature | Official vendor page, dated |
| Measured performance | Reproducible direct test |
| User sentiment | Multiple recent community sources; label as anecdotal |
| Commercial terms | Current affiliate/network terms |
| “Best” / recommendation | Transparent decision criteria and alternatives |
If a claim cannot be placed in the ledger, either verify it or remove it.
7. Render multiple formats from one evidence object
The same verified record can generate:
- a high-intent owned article;
- a 20–40 second short-form demo;
- a comparison table;
- a newsletter note;
- a longer tutorial script.
This is the safe way to “scale content”: scale the reuse of verified evidence, not the number of unsupported claims.
8. Human QA should have a checklist
- Did we actually use the product when the article says we did?
- Are pricing and availability dated?
- Could any statement be interpreted as a fabricated testimonial or earnings promise?
- Are affiliate relationships disclosed?
- Does the article include a real limitation or better-fit alternative?
- Do screenshots expose private data?
- Is the CTA consistent with the evidence?
9. Measure the funnel, not vanity reach
Views are an attention metric, not a business result. Track the chain:
impression/view → retained attention → profile/page visit → outbound affiliate click → merchant conversion → revenue → refund/churn where relevant
For owned content, useful secondary signals include search impressions, scroll depth, return visits and email opt-ins. Revenue should ultimately reconcile to the affiliate network dashboard.
10. Make the system able to kill ideas
Automation that only produces more is incomplete. Add rules for stopping weak experiments: repeated poor retention across different hooks, high clicks with no conversions, product quality deterioration, bad economics, or excessive production effort.
A system that kills bad ideas quickly creates more value than one that generates 100 drafts per day.