The Architecture of Isolated Agents Protects AI Content from Falling into SERP Consensus
AI content rewriting suffers from a known bug: the model invents https://x.com/shannholmberg/status/2046891905389334958 missing pieces, deviates from the original vector with each iteration, and cleans up specific sentences that provided positions.
`Autoreason` addresses this vulnerability by cutting visibility between roles — each agent sees exactly what it needs, allowing the original strength of the text to survive.
Agent Architecture
- Active Champion: the current page (header, subheading, H2 structure, block order).
- Critic: sees the current page + target keyword + top 3 competitors from SERP + `ICP` language from sales calls.
- Provides only critique (for example, “H2 on line 142 rephrases a non-entity, the example in block 3 does not match the ICP lexicon from transcripts”).
No Rewriting.
- Author `B`: sees the brief + original + critique.
- Generates a corrected version.
- Synthesizer: sees the original + version `B` in random order.
- Combines the strongest pieces from both.
- Judging Panel: 3 fresh agents rank `A`, `B`, and `AB` using the Borda method (blind voting).
- Convergence at `streak=2` (prevents infinite loops).
The knowledge base grounds the result in business realities instead of template `SEO` advice:
- GSC: queries that generate impressions but lose `CTR`, keywords with zero optimization, pages with month-to-month drops.
- Competitor data (Ahrefs/Moz): interlinking patterns of top 3 competitors, anchors, entity gaps compared to your site.
- GA4 conversion data: judges evaluate options based on whether “ICP will convert after reading,” not “does it look prettier.”
- Transcripts of sales calls and support tickets: agents write in the language of clients, not in marketing jargon.
- High-performing pages: the synthesizer embeds working patterns of a specific domain, not consensus from the results.
Use cases follow the same cycle with context-specific judges:
- Service pages: the judging panel evaluates based on the criterion “will the lead forward this within the company,” using the lexicon of won deals from `CRM` as a benchmark.
- Free tool pages: the judging panel compares with queries that generate impressions in `GSC`, and how competing tools structure blocks above the fold.
- SEO software templates: run the template through the cycle itself (not each page).
- Judges evaluate a sample of 20 filled pages for entity coverage, uniqueness, and conversion potential.
Insights
— The problem of pollution (when AI quietly cleans specific, hyper-relevant sentences, deeming them redundant) is addressed through strict limits on editing instead of complete rewriting — tighten the `title/H1` to fit the intent, reassemble the intro to capture the snippet, add entity/FAQ blocks for `GEO`, refresh internal linking; do not request a rewrite of pieces that are already ranking.
