Reporting cadence
| Cycle | What you do |
|---|---|
| Weekly (internal) | gsc_pull deltas by page group; investigate any anomaly the same day; log in SEO OS. Friday: Console “Copy status” into Slack + JSON export. |
| Monthly (client) | Looker Studio dashboard (GSC + GA4) + written narrative: outcomes vs forecast → what we did → decisions taken → next decisions → risks. Every page cohort gets a verdict: Grow / Defend / Improve / Merge / Repurpose / Noindex / Remove. |
| Quarterly | Re-pull the 16-month baseline; forecast-vs-actual calibration; re-run project classification; review refresh triggers (comparable decline, factual change, SERP shift — never “the page is old”). |
Forecast model
demand × target CTR × rank probability × conversion rate × qualification × close rate × value, ramped over months, confidence-banded, minus cost. Blocked until every input has a source and a date — no silent defaults, no exceptions. Estimates are always labeled as estimates.
Strategy engine — 7 steps
- Economics first: contribution margin or lead value → value per organic session → break-even effort threshold. No number, no strategy.
- Demand truth: GSC before keyword tools — brand vs non-brand, striking distance (positions 4–15), CTR gaps.
- Market evidence: live SERPs per cluster (logged country / device / date); page-level competitor teardowns; define the evidence advantage — what this client can prove that ranking pages cannot.
- Intent-fit: one primary intent = one URL; resolve cannibalization by merge / redirect / reposition before creating anything new.
- Sequencing: capture existing intent before creating demand — fixes → striking distance → bottom-funnel → supporting → GEO → authority.
- Forecast + human verdict: Pursue now / Pursue after support / Long-term / Reject.
- Kill / Iterate / Scale: no verdict before 30 days AND a minimum-impressions threshold; monthly cohort decisions; quarterly calibration; immediate stop only for ethics or policy risk.
The 12 QC rules — non-negotiable
- No access, no audit — ever.
- Every finding: affected URLs + evidence reference + date + acceptance test.
- Sample-verify 10 items or 10% of any tool or AI output before a client sees it.
- Publish gate per URL: index / canonical · rendering · schema valid · analytics fires · CTA works. Any fail = no publish.
- Executor and reviewer are different people for P0 work, deploys and YMYL. Solo adaptation: Claude adversarial review + 24h cooling-off self-review.
- No ranking or traffic guarantees; forecasts only from complete inputs; estimates labeled as estimates.
- Link ethics: no purchased links, PBNs or expired-domain redirects; disavow only for documented risk.
- SERP evidence older than 30 days is expired for decisions.
- AI drafts always pass human fact / SME review; no mass generation; YMYL adds credentialed sign-off.
- Claude never self-authorizes deployment: branch → PR → tests → rollback documented.
- Client data never enters shared or group-buy tool environments.
- Never blend a recovery or migration event with unrelated changes — one major variable at a time.
Default answer: if a tactic is not in this system, the answer is no until it passes these rules.