SEO·DECAY / SYSTEM ARCHITECTURE
4 AGENTS · 2 OPERATIONAL
FOUR AGENTS.
TWO ARE REAL. ON PURPOSE.
01 DECAY + VISIBILITY AUDITOR
OPERATIONAL

Per-page 6-month traffic decay (GSC — simulated for third-party domains), AI crawler flags (simulated), 64 fixed prompts through live Tavily retrieval + LLM synthesis. Grounded answers serve as the AI-Overview proxy.

tools/run_agent1.py
02 ROOT CAUSE
OPERATIONAL

Classifies each flagged page among exactly four causes — content gap, AI-answer absorption, core algorithm update, intent shift — with reasoning over the evidence bundle. Never "competitor has more content."

tools/run_agent2.py
03 UPDATER
DIAGRAM ONLY

Generates the rewrite brief: which subtopics to add, which queries to concede, which to fight. Human review before anything ships.

NO CODE — CONCEPT ONLY
FROM AGENT 02 CONTENT GAP AI ABSORPTION CORE UPDATE INTENT SHIFT VERDICTS + EVIDENCE ADD MISSING SUBTOPICS CONTENT GAP — FIGHT IT CONCEDE THE QUERY AI ABSORPTION — DON'T REWRITE HOLD FOR REVIEW UPDATE OR INTENT — WAIT HUMAN REVIEW GATE NO SHIP WITHOUT HUMAN OK APPROVED BRIEF TO AGENT 04 AUTO-REWRITE NOT BUILT DASHED = DELIBERATELY NOT BUILT · CONCEPT ONLY — NO LIVE CALLS

Reads Agent 2's root-cause verdicts and turns them into a rewrite brief. A content gap means adding the missing subtopics and fighting. AI-answer absorption means conceding the query — no rewrite fixes it. Algorithm updates or intent shifts mean hold. Every brief stops at a human gate; nothing is auto-written.

04 PUBLISHER
DIAGRAM ONLY

Drafts to CMS via API, before/after traffic and citation tracking wired back to Agent 1 to close the loop.

NO CODE — CONCEPT ONLY
APPROVED BRIEF FROM AGENT 03 HUMAN-EDITED DRAFT WRITER WORKS BRIEF CMS API PUSH WORDPRESS / WEBFLOW DRAFT + MEDIA + META LIVE PAGE PUBLISHED URL RE-AUDIT GSC BEFORE-AFTER CITATION RE-CHECK BACK TO AGENT 01 NEXT BATCH INPUT AUTO-PUBLISH NOT BUILT DASHED = DELIBERATELY NOT BUILT · CONCEPT ONLY — NO LIVE CALLS

Takes the human-approved brief, hands it to a writer, pushes the finished draft to the CMS via API, then measures — before/after GSC traffic and a citation re-check on the same fixed prompts. Results feed Agent 1's next batch, closing the loop. A human clicks ship; nothing is auto-published.

COST MODEL
HOSTING$0/MOSTATIC FILES, NO SERVER PROCESS EXISTS
RUNTIME API CALLS$0SITE READS JSON AT BUILD TIME. NO KEY SHIPS.
AUDIT BATCH$0.00~55 TAVILY (FREE 1,000/MO) + ~70 LLM CALLS (FREE TIER), RUN ONCE BY HAND
RUNAWAY SPENDNONEEVERY CALL PASSES A HARD BUDGET GUARD THAT ABORTS PAST FIXED CAPS
DELIBERATELY OUT OF SCOPE

· Auto-rewriting or auto-publishing (Agents 3–4 stay diagrams until verifiable)

· llms.txt generation — no engine has confirmed consuming it

· AI-Overview SERP scraping — the LLM-over-Tavily synthesized answer is the proxy, labeled as such

· Any scheduled/cron execution — the batch is workflow_dispatch only