Phase 3: Synthesis Process

Running all skills generates a large volume of recommendations. The synthesis step is where the audit becomes useful.

Before synthesis: run trend-scan

Use trend-scan on the Google Search Console CSV exports collected in Phase 1. A Dates export establishes traffic direction; Pages or Queries exports show concentration, and comparable prior/recent exports identify declining or missing pages and queries. See the GitHub skill instructions for the helper commands and supported inputs.

Save the output as skills/trend-scan.md in the client audit folder. Use the recent trend alongside the longer-period total when deciding what leads the report. Investigate each flagged mover against the change timeline, technical evidence, and live search results; a detected decline does not establish its cause.

If history is missing, incomplete, or not comparable, record the limitation explicitly. Do not infer a healthy trend from a large aggregate total. Install the public helpers if the skill is unavailable.

Step 1: Collect all skill outputs into one document

Paste each skill’s output into a staging document with a clear heading for each skill. Don’t edit yet — just collect.

Step 2: Deduplicate

Multiple skills will surface the same finding (e.g., comparison pages appear in /seo-audit, /programmatic-seo, and /competitors). Keep the most detailed version; strip the others.

Step 3: Reconcile the change timeline

Before diagnosing a traffic or ranking mover, align it with change-events.md: deployments, migrations, template/plugin releases, content and analytics changes, GSC Manual Actions/Security Issues, Google Search Status incidents, and data anomalies. Keep preliminary 24-hour data separate from finalized history.

A date match is correlation. Write “consistent with” unless reproduction, rollback, logs, or another controlled check confirms the mechanism. Do not prescribe a seven-day ranking-only auto-rollback; roll back for failed technical acceptance criteria or pre-agreed business guardrails, otherwise measure at 2/4/8 weeks using an unaffected comparison where possible.

Step 4: Group by shared root cause

Map related finding IDs to their common template, component, plugin, CMS field, or configuration before creating recommendations. Record the affected URL pattern/count, representative failing URL, proposed single fix, unaffected regression sample, and confidence. One shared mechanism becomes one recommendation even when it affects hundreds of pages. Keep page-by-page actions only for genuinely isolated faults.

Step 5: Reconcile content lifecycle decisions

Join the URL register with comparable search periods, conversions, backlinks, user/business value, query→Pages, historical SERP, and technical evidence. Every affected URL gets one provisional decision: Protect, Refresh, Merge, Differentiate, Redirect, Noindex, Remove, or Investigate. Destructive routes remain Red/Human decision and cannot be based on age or traffic alone.

Step 6: Compile evidence-backed content briefs

Every proposed new, refresh, or merge action gets a brief before entering the 90-day plan: persona/task, conversion goal, demand source/date/market, current Google landing URL(s), live SERP intent/page type, unique first-party contribution, outline, CTA, links, supported schema, missing facts/approvals, acceptance, baseline, and 2/4/8-week measurement. AI may draft after approval but cannot invent claims, experience, evidence, statistics, or competitor facts.

Step 7: Cross-reference with raw data

Every recommendation needs to be grounded in the client’s actual data:

  • “Build comparison pages” is generic. “Build /compare/golive-vs-enov8/ because Enov8 holds position 1-2 on the head term and the client is at position 17” is specific and defensible.
  • Check each skill recommendation against the GSC, GA4, PSI, and Squirrel Scan data you collected.

Step 8: Assign tiers

  • Tier 1 (Week 1): Critical indexing bugs, any 404 bugs fixable in under 1h, broken schema on key pages, sitemap/robots breakage. (Security findings do NOT belong in Tier 1 — see the Security & Platform Risk section.)
  • Tier 2 (Month 1): Highest ROI items — Marketplace/conversion optimization, performance, highest-impression content upgrades, tracking gaps
  • Tier 3 (This quarter): Content cluster, comparison pages, pSEO launch
  • Phase 2 (Defer): Items needing external data not yet collected (backlinks, log files, competitor gap analysis)

Step 9: Assign impact + effort

For each item:

  • Impact 1-5: How directly does this item affect organic traffic, rankings, or conversions?
  • Effort: T-shirt size (15min / 30min / 1h / half-day / 1 week / quarter+)

No owner field. The tool doesn’t know the client’s team or resources, so don’t invent one — effort size already signals whether it’s a self-serve fix or a developer task; let the reader assign their own people.

Step 10: Calculate the SEO Health Score

Before writing the 90-day plan, calculate the 0–100 SEO Health Score. The AI assigns a 0–100 score to each of 7 B2B SaaS-weighted categories based on the collected data (Content Coverage 22%, Technical SEO 20%, On-Page 18%, Performance 12%, Schema Markup 10%, AI Visibility 10%, Programmatic SEO 8%). AI Visibility must use the fixed prompt/citation benchmark and observed third-party citation evidence; do not award points merely for llms.txt or unverified “AI-specific” markup. Apply the weights and produce the per-category breakdown table. The score identifies the highest-impact levers, which should shape 90-day plan prioritization. Severity tiers: Critical (0–39), High (40–59), Medium (60–79), Low (80–100).

Step 11: Write the 90-day execution plan first

Before detailing individual recommendations, write the 90-day plan section. This forces you to decide what actually matters vs. what’s just interesting to include. If an item doesn’t make it into the 90-day plan, ask why it’s in the report at all.

Step 12: Answer the Acquisition Channel Questions — last, on purpose

The report closes with a Q&A section framed by one goal: turn the website into a predictable acquisition channel. The standing questions (runbook §5.7) cover website structure, conversion optimization, product pages, landing pages, lead magnets, SEO opportunities, and content strategy — plus two focus questions: what content attracts buyers vs. just traffic, and how blog → lead magnet → demo/trial should connect. The section ends with three deliverables: a website improvement roadmap, a lead magnet plan, and content priorities for the next 6 months.

Why last: none of these is answerable from a single source. “Buyers vs. traffic” needs GSC and GA4 conversion paths; the lead-magnet plan needs the persona, the conversion architecture, and demand data; the 6-month priorities need the tiers and 90-day plan to already exist so they can point into them instead of duplicating them. The AI answers every question itself, citing collected data — where data is missing (no conversion tracking, say), the honest answer is the gap plus its Tier 1 setup action.

Before delivery: run claim-check

Run claim-check on <client-slug>-seo-audit-YYYY-MM-DD.md. Resolve each flagged time-sensitive claim against a primary source: confirm it and record the source and verification date, correct it, or remove it. Save the verification notes as skills/claim-check.md in the client audit folder.

The script finds potentially stale or undated statements; the agent performs the source verification. A clean scan does not prove that the report is accurate. Complete the report QA checks before generating the action runbook.