SEO Improvement Plan v2 — Closed-Loop AI/GEO Automation

Date: 2026-04-28

Owner: Cinder (autonomous after Costa greenlight)

Goal: Continuously improve organic + AI-Overview citations across all 29 sites without Costa in the loop.

What we have today (running)

Cron	What it does	Action surface
`gsc-weekly-pull` (Mon 11:00)	Pulls 28d clicks/impressions/CTR/position per site	JSON + Telegram digest
`gsc-striking-distance` (daily)	Identifies queries in pos 4-20 (one-content-push-from-ranking) + risers/fallers	JSON + Telegram digest
`ga4-feedback-loop` (daily)	Sessions + conversions per site	JSON + Telegram digest
`thin-content-dryrun` (daily)	Identifies pages under 700 words	DRY-RUN ONLY — no fixes
`indexnow-daily` (daily)	Pings IndexNow for crawl signals	27/29 succeed (Orlando + Bloomington 403)
`innovation-sprint` (Mon 06:00)	Weekly research digest covering Google updates, AI Overview shifts, R&R industry, competitor moves	Telegram

The gap: Every loop is one-way. Signal → report → done. Nothing closes back to action.

What's missing — close-the-loop

Loop 1 — Striking-distance fallers → page refresh

Trigger: Query falls 3+ positions in gsc-striking-distance daily run.

Action: Enqueue (domain, page, query, current_pos, prior_pos) to data/action-queue/seo-refresh.jsonl.

Worker (tools/seo-refresh-worker.py, runs nightly 02:30):

1. Pulls top item from queue
2. Reads target page, identifies query intent + content gap (Sonnet via Claude API)
3. Drafts a 200-400 word content block that answers the query directly
4. Inserts block via Edit tool
5. Triggers IndexNow ping
6. Logs to data/action-queue/seo-refresh-completed.jsonl
7. Telegram morning brief surfaces what shipped

Loop 2 — Thin-content auto-expand

Trigger: Page identified by thin-content-dryrun AND has GSC impressions but 0 clicks (= ranking but not converting).

Action: Same as Loop 1, except prompt focuses on E-E-A-T expansion (case studies, FAQ, pricing transparency).

Safety rails:

- Whitelist content types: services/, locations/, blog/*. Never touch contact, privacy, terms.
- 1 page per site per day max.
- Hard daily cap: 5 pages total across portfolio.
- Costa-greenlight required for first 5 production runs (Telegram: "approved" / "reject").

Loop 3 — AI Overview / GEO citation tracking

Trigger: Weekly innovation-sprint pulls competitor SERP samples + AI Overview snippets for top-10 queries per site.

Action:

1. Detect AI Overview citations (us vs. competitors)
2. If site is mentioned in AI Overview: log win to data/geo-wins.jsonl
3. If competitor is mentioned and we're not: queue the relevant page for GEO-optimization rewrite (uses geo-optimize skill)
4. GEO-optimize prompt focuses on: structured data (FAQ, HowTo, Product schema), question-answer format, clear factual claims with citations, conversational language matching how AI synthesizes answers

Schema deployment:

- Auto-add FAQPage schema to any page with H2/H3 question patterns
- Auto-add LocalBusiness schema to all locations/* pages
- Auto-add Service schema to all services/* pages

Loop 4 — Internal-link sweep

Trigger: Page with rising impressions but flat clicks (CTR drop).

Action: Auto-add 2-3 contextual internal links FROM higher-authority pages on the same site TO the rising page. Concentrates link equity.

Loop 5 — IndexNow + GSC URL Inspection (the Costa-asked-for relay)

Trigger: Any of Loop 1/2/3/4 ships a page edit.

Action:

1. Ping IndexNow daemon (existing)
2. Queue URL to data/gsc-submit-queue.jsonl for Playwright + Chrome user-data-dir relay (per docs/gsc-relay-options-2026-04-28.md, Option 1)
3. GSC relay daemon (Mac mini, 03:00 CT) processes queue, submits via URL Inspection, screenshots verification

AI/GEO-specific layer

GEO = Generative Engine Optimization (ChatGPT, Perplexity, Claude, Google AI Overviews).

What changes for AI vs traditional SEO

Traditional SEO	GEO addition
Keywords in title/H1	Q&A blocks matching how users ask AI assistants
Backlinks for authority	Mentions on authoritative third-party content (Wikipedia, industry directories, .gov, .edu)
Schema for rich snippets	Schema for fact-extraction (Product, Service, FAQ, HowTo, BreadcrumbList)
Page speed	Crawlable structured facts above the fold
Title CTR	Direct answer in first 60 words

Auto-implementation

Each site gets a per-page "AI-readability score" computed weekly:

- 30 pts — has FAQ schema with Q matching top GSC query
- 25 pts — answers the primary intent in first 60 words
- 20 pts — structured data validates against schema.org
- 15 pts — has at least one citation/source link
- 10 pts — content updated within last 60 days

Below 60/100: auto-queued for GEO-optimization rewrite via Loop 3.

AI-detection prompt (runs weekly per site)

For each site, query Perplexity/Claude API with:

- "What is the best [niche] in [city]?"
- "Who do you recommend for [service] near [city]?"
- "Compare [our brand] vs [competitor]"

Score: are we mentioned? In what position? With what description?

Track in data/geo-citations.jsonl over time. Trend = our GEO trajectory.

Per-site dashboard (single source of truth)

Build tools/site-health-dashboard.py (cron daily 06:30):

For each of 29 sites, write data/site-health/<domain>.json:

{
  "domain": "elkhornhardwood.com",
  "as_of": "2026-04-28",
  "gsc_28d": { "clicks": 13, "impressions": 985, "ctr": 1.3, "position": 17.3 },
  "ga4_7d": { "sessions": 24, "conversions": 1 },
  "twilio_7d": { "real_calls": 0, "spam_calls": 0 },
  "thin_pages": 3,
  "ai_readability": 52,
  "geo_citations_30d": 0,
  "queued_actions": [
    {"type": "seo-refresh", "page": "services/refinishing", "reason": "fall -3.4 pos"},
    {"type": "thin-expand", "page": "locations/papillion", "current_words": 312}
  ],
  "trend": "stable"
}

Surfaced in morning brief as a single 5-row exception report:

> Site health alerts (3):

> - elkhornhardwood.com — 0 calls 7d, 24 GA4 sessions, 2 actions queued for tonight

> - phxpoolresurfacing.com — fallers detected on 3 queries, refresh queued

> - orlandoconcretedriveway.com — IndexNow 403 7 days running, key needs regenerated

Implementation timeline (autonomous, no Costa input needed)

Tonight (overnight-v7)	Tomorrow	Day 3	Week 2
Build action-queue/seo-refresh.jsonl + worker	Wire striking-distance → queue (Loop 1)	Wire thin-content live mode (Loop 2) on services/* whitelist	Loop 3 (AI Overview citation tracking)
Fix IndexNow 403s on Orlando + Bloomington	Build site-health-dashboard.py + cron	First 5 Loop 1/2 runs Costa-reviewed	Loop 4 + Loop 5 wiring
Patch 4 wrong-tel: numbers (DONE earlier today)	Greenlight Loop 1 production	Greenlight Loop 2 production	GEO citation tracker live

Costa input required ONLY at:

1. First 5 Loop 1/2 runs — verify auto-rewrites are quality (Telegram: "approved" / "reject")
2. Confirming dedicated ITD Chrome profile for GSC relay daemon (per gsc-relay-options-2026-04-28.md)
3. Sourcerow domain + budget approval (separate ticket, not SEO)

After greenlight, all 5 loops run autonomously. Telegram morning brief surfaces exceptions only.

Why this works (ranked by impact)

1. Loop 1 (striking-distance refresh) — biggest mover. ~10-20 queries per site sit at pos 4-20. Each one closing into top-3 = 5-10x the click volume. With 29 sites × 10-20 queries = 300-600 active opportunities, automated refresh of even 20% = 60-120 ranking lifts in 90 days.

2. Loop 3 (GEO citations) — biggest 12-month opportunity. As AI Overviews replace traditional SERPs (already 30% of US queries by Apr 2026), being cited by AI = the new SEO. Most R&R competitors aren't optimizing for this yet.

3. Loop 2 (thin-content) — quick wins. ~7 thin pages per site × 29 sites = ~200 pages that could get to a baseline. Even +50 sessions/page/month = significant aggregate.

4. Loop 4 (internal links) — compounding effect on Loop 1.

5. Loop 5 (IndexNow + GSC relay) — accelerator for all 4 above. Without it, edits take 2-7 days to reflect in rankings; with it, often 24-48 hr.

Costa's role (what I need from you)

Once:

- Greenlight the first 5 Loop 1+2 production runs (review the auto-rewrites in Telegram)
- Pick: dedicated ITD Chrome profile OR daily-driver for GSC relay

Ongoing (weekly, ~5 min):

- Skim morning brief site-health alerts for sanity check
- Reject any auto-action if it doesn't fit your judgment

That's it. Everything else is mine.