The closed loop, in depth.

01 · Measure

For each brand we run a representative set of buyer-intent prompts ("best issue tracker for engineering teams", "alternatives to Asana for SaaS", etc.) against each enabled engine. Every prompt is run N times (default 3–5) to get a statistically meaningful sample on each engine independently.

For each probe we record three signals:

Retrieved

Did your domain appear anywhere in the engine's live search/retrieval set?

Cited

Did the engine actually link to your domain in its answer or Sources list?

Mentioned

Did the engine name your brand in the prose, even without a link?

Rates are reported as proportions with Wilson 95% confidence intervals. No "scores", no synthetic prompts, no rankings implying false precision.

02 · Diagnose

Rule-based gap detection runs on every audit:

• Low retrieval — your own domain rarely surfaces in live search
• Mention-without-citation — engines vouch via third parties, not you
• Weak prompts — specific queries where the brand underperforms its average
• Citation targets — high-authority non-competitor domains AI trusts in your category
• Worldview errors — factual inaccuracies, missing facts, stale framing
• Per-engine disparities — when engines disagree by ≥ 30 pts, the play is platform-specific

Each gap is grounded in a measured metric and rolled into an LLM-written remediation plan.

03 · Generate

For each diagnosed gap we produce concrete artifacts:

• A publish-ready content brief using the only causally-evidenced GEO levers (statistics, quotations, cited sources — see Methodology)
• FAQ schema.org JSON-LD derived from real buyer Q&A
• Outreach target list — the specific third-party domains AI already trusts in your category, with the specific angle for each
• Worldview corrections — where to fix factual errors so AI engines pick up the correction (Wikipedia, Wikidata, Crunchbase, owned site)

04 · Publish

The artifact compiler turns the content brief into a publish-ready file (Markdown, MDX, or HTML) with proper frontmatter, canonical URL, and embedded JSON-LD. The Git connector pushes it straight to your content repository — local commit or pull request, your choice. Idempotent on identical content. Full provenance recorded.

Sample output frontmatter:

---
title: "Why Engineering Teams Are Switching to Linear..."
slug: "best-project-management-tool-software-development-teams"
canonical: "https://linear.app/best-project-management-tool-..."
source_run_id: "linear-20260613T223433Z"
published: 2026-06-14
---

05 · Re-test

Every audit is stored as a run in SQLite. Re-auditing after a publish produces a new run, and the compare module computes per-metric deltas across runs. A change is flagged significant only when the Wilson 95% CIs don't overlap — so we don't claim lift on what's really sampling noise.