SEO

The SEO loop audits one website and gives you a 0–100 SEO Health Score plus a prioritized action plan. It never changes your site — it observes and recommends, filing each issue worth acting on as a ticket on your Tasks board.

What it checks

The audit runs in two lanes:

• Mechanical checks — title, meta description, canonical, html lang, viewport, Open Graph, JSON-LD, h1, image alt, robots.txt, sitemap, Core Web Vitals, and organic-traffic regression. These are rule-based, deduped, and auto-close themselves when a fix lands.
• Judgment checks — the things a regex can't compute: content depth and topical authority, E-E-A-T signals, internal-linking strategy, keyword and topic gaps, AI-search readiness (GEO / llms.txt / citability), and a client-side-rendering check that compares your JS-rendered page against the raw server HTML a crawler actually indexes.

Core Web Vitals come from Google PageSpeed Insights lab metrics (LCP, FCP, TBT, CLS) per strategy (mobile / desktop) — synthetic lab numbers, not real-user field data. Every finding rests on something a tool actually returned; if a signal can't be observed, the audit says "not measured" rather than guessing.

Prerequisites

• A product with a website URL set. Without one (and no Search Console property to fall back on), the loop skips.
• Optional but recommended: connect Google Search Console. Without it, the audit still runs on technical and on-page signals; with it, you also get real clicks, impressions, CTR, position, and your top queries and pages folded into the score.

Run it

Ask the dispatcher in plain language — for example:

audit my SEO

It also runs on its own roughly every 30 days. On an autonomous wake it will skip if a full audit already ran in the last ~25 days and nothing material changed; a user-triggered "audit my SEO" always runs.

What you get

A report artifact with the 0–100 score and the narrative, plus a ticket for every actionable issue:

• Mechanical issues are filed automatically and carry a stable fingerprint, so re-runs dedupe them and they close themselves once fixed.
• Judgment issues are filed with a severity and category for you to triage.

Approving an audit can spawn a coding task that applies the safe, mechanical fixes as a pull request — so the highest-leverage technical fixes can go from "found" to "PR" without you writing them.

Notes & limits

• Recommend-only. Nothing on your site is changed by the audit itself.
• Lab Core Web Vitals are synthetic; treat them as directional, not field truth.
• Findings are only as deep as the tools can see — a JSON-LD block injected client-side by a tag manager won't appear in the server HTML the audit reads, and it's reported as "absent in server HTML" rather than missing.
• Set how often the audit runs from Cadence.