Human Strategy. Agent Execution. The Site That Optimizes Itself the Right Way.

The Problem

SEO and AEO work is slow, manual, and easy to drop. Keyword research, content-decay analysis, striking-distance opportunity identification, on-page fixes, and Answer Engine Optimization updates all compete for time that never shows up consistently.

SEO compounds. Every week without optimization is a week of compounding authority the site doesn't earn. A human-dependent process that runs inconsistently produces inconsistent results, which is indistinguishable from no process at all.

But the inverse is also a problem. A fully autonomous system that makes strategic decisions without human review compounds in the wrong direction. The content that earns rankings and citations is the content with practitioner voice, real opinions, real client data, real domain expertise. An agent left to its own strategic devices produces marketing-grade beige content that nobody cites and nothing ranks for. The challenge is dividing the labor between human strategy and automated execution in a way that compounds steadily without compounding into the wrong work.

What We Built

We built a dual-MCP engine that exposes 21 SEO and AEO tools to a Claude agent running inside Cowork, then layered a human-strategy plus agent-execution split on top of the tooling.

The human input layer is Joseph. Joseph sets cluster strategy (which topic clusters to invest in), decides which verticals are co-equal #1 priorities, runs the Strategic Assumption Registry that says what bets the program is making and when each assumption gets re-tested, gives voice red-lines on every draft (the first Cowork pillar shipped went through a v1-to-v2 rewrite because the v1 voice was too technical for the dual-persona target audience), and makes the publish go-or-no-go call on every piece that crosses the 30%-structural-change threshold.

The agent execution layer is research, analysis, drafting, and cadence execution. The agent does not make strategic decisions. It does not decide which queries to target, which verticals to invest in, when to graduate a piece from draft to publish, or how to sound. It runs the system Joseph designed, surfaces what the data says, drafts what the strategy says to draft, and waits at the red-line. Below the 30%-structural-change boundary, the agent ships directly via the Content MCP and triggers a Vercel deploy. Above the boundary, the agent queues a draft and a preview and waits for Joseph.

Joseph's actual time investment is roughly 2 to 4 hours per week on the program, almost all of it spent on strategic decisions and voice review. The agent's running time is roughly 10 to 15 hours per week of compute, doing research, drafting, and analysis that would otherwise consume 20 to 30 hours of an SEO operator's time. The leverage ratio is high because the work that compounds (the strategic decisions and the voice integrity) stays human.

“The agent runs the system. The strategy stays human. SEO that compounds in the right direction, not just any direction.”

The System Architecture

Two MCP (Model Context Protocol) connectors expose SEO and AEO data to the agent. DataForSEO contributes 12 tools covering keyword research, SERP analysis, competitor research, and on-page audits. Google Search Console contributes 9 tools covering indexing inspection, striking-distance keyword detection, content-decay scanning, cannibalization checks, and click-through-rate analysis. A third MCP, Automaton's own Content MCP, gives the agent read and write access to the live site, including pages, blog posts, image generation, and voice-clone tooling.The agent runs inside Claude Cowork on a published cadence. Daily five-minute health checks Monday through Friday scan algorithm news, run a rotating two-query AI-citation check, diff the live site against the inventory, push any auto-tier fixes that are queued, and post a short Slack summary. Themed working sessions happen 3 to 4 times per week: Mondays for vertical-pillar work, Wednesdays for build-vs-buy or Cowork-cluster work, Fridays for infrastructure work. Each working session is constrained to ship one deliverable. Bi-weekly, the agent runs a full strategic review: ten-query citation retest, GSC analytics deep-dive, brief generation, strategic plan refresh.The output is two things. Direct content updates via the Content MCP for auto-tier work (small structural changes, internal-link sweeps, freshness date bumps). Prioritized action briefs to Slack for approval-tier work (new pieces, major edits, schema changes, anything over a 30%-structural-change threshold). A separate scheduled meta-audit runs bi-weekly to examine the agent's own work product and ask whether the work it's producing is still the right work.

The Results

Thirty days into the publishing cadence: 9 pillar-grade pieces shipped, 8 of 9 indexed by Google, 28-day organic impressions grew from under 50 to over 3,500 (a 4.6x increase in the seventh week alone).

Three of the program's ten target queries for Answer Engine Optimization are now cited multiple times in Google's AI Overviews. The strongest single query has 8+ inline Automaton citations across six AIO body sections. A supplementary 'AI automation ROI' query flipped from zero citations plus a Featured Snippet to 5+ AIO citations with the entire 'Where to Focus' list lifted verbatim from the piece, all in a single weekly cycle.

Joseph's actual time on the program runs 2 to 4 hours per week. The agent's compute time runs 10 to 15 hours per week. Comparable manual SEO operator time would run 20 to 30 hours per week. The leverage ratio sits around 7 to 10x.

The compounding effect: every week the engine runs, the site gets marginally better. The compounding effect of the human strategy layer is the part most autonomous-SEO claims undersell. Every week's strategic decisions, voice red-lines, and red-line learning compound into a system that increasingly produces the right work, not just more work. The field-report blog post at /insights/ai-overview-citation-30-day-field-report is the public-facing view of the same data.