diff --git a/docs/operations/visibility-stack-roadmap.md b/docs/operations/visibility-stack-roadmap.md
new file mode 100644
index 0000000..474c197
--- /dev/null
+++ b/docs/operations/visibility-stack-roadmap.md
@@ -0,0 +1,73 @@
+# Visibility Stack Roadmap — SEO / AEO / GEO Integration
+**Status:** Frame LOCKED · 2026-08-16 · Signal source: Mohart framework post (x.com/natan_mohart/status/2088994329272652168) + concurrent llms.txt / agent-readable-data discourse
+**Rule of this doc:** every phase has a hard trigger. No phase starts on vibes.
+
+## Positioning (the sell)
+VeriPath keeps a business's verified facts consistent across every surface — so both
+search engines and AI systems (ChatGPT, Gemini, Claude, Perplexity) represent the
+business accurately.
+- **SEO** — found by search engines (GBP, Bing, listings) — *built and in production.*
+- **AEO** — chosen as the direct answer (structured data, JSON-LD, NAP) — *partially built; website scope is client-negotiated per DR-001.*
+- **GEO** — recommended/cited by AI systems (consistent, machine-digestible verified truth) — *the canonical record IS the foundation; no GEO product exists yet, by design.*
+
+Owner-facing one-liner (Ashley call and first 2–3 clients):
+> "We keep your facts consistent so the systems that recommend you have the right information."
+
+Acronym framing (SEO/AEO/GEO) is for partners, chambers, and later-stage sales —
+never the first thing an owner hears.
+
+## Phase 0 — NOW (in progress, zero code)
+**Trigger:** none — active.
+- Three-layer language in calls and sales conversations (above).
+- Trust receipt carries one paragraph: surfaces kept consistent for both search
+  engines and AI recommendation. (Prose artifact — does not touch the locked
+  beta report format.)
+- Canonical record remains the single source of truth; all surfaces measured against it.
+**Exit:** first trust receipt sent.
+
+## Phase 1 — After first loop closed
+**Trigger:** Phoenix loop complete (onboarding → Lovable fix → re-run → trust
+receipt sent → monitoring day 0).
+- **P1a.** Report format bump v1.1: tag high-severity consistency findings
+  (hours / schema / NAP) as affecting both search and AI recommendation.
+  First change to the locked beta format — deliberate, post-proof.
+- **P1b.** Decision: lightweight GEO check enters the standard receipt or not.
+  Test = fixed local prompt set (e.g., "best salon in Cameron Park"), manual or
+  scripted verification, against the canonical record. Decide from real owner
+  feedback + whether the check produced anything the owner cared about.
+**Exit:** format v1.1 committed; P1b decision recorded (DR).
+
+## Phase 2 — After 2–3 clients onboarded, monitoring running
+**Trigger:** ≥3 clients with active monitoring clocks.
+- llms.txt / website structured data as a **client-negotiated website scope**
+  (never silent website access — DR-001 holds).
+- If P1b check proved signal: standardize it in the receipt and monitoring
+  cadence (DR-002 mechanism absorbs it).
+- Entity-clarity pass in intake: name/brand variants, disambiguation fields,
+  only if real clients showed the problem.
+**Exit:** one client with a website-scope deliverable shipped (or explicit
+evidence it's not wanted → document, move on).
+
+## Phase 3 — After unit economics proven
+**Trigger:** 8–12 clients, monitoring at minimal daily attention, proven
+per-client margin.
+- Productize the visibility layer as its own tier (the "recommendable by AI"
+  offering) with its own pricing and the GEO check as the measurable artifact.
+- Partner/chamber channel uses full SEO/AEO/GEO framing.
+
+## Standing rules (binding, all phases)
+1. **Verified truth before visibility.** No visibility artifact ships before the
+   client's canonical record exists and is owner-verified.
+2. **Canonical record is the single source of truth.** Anything agent-readable
+   (llms.txt, schema, GEO check output) is a *rendering* of it — never a new
+   source. Any proposed artifact answers yes/no: does it render verified truth,
+   or create a second truth?
+3. **No second source, no new surface without a client who asked for it.**
+4. DR-001 (website scope) and DR-002 (cadence) stand unchanged unless explicitly superseded.
+
+## Non-goals (until Phase 1 trigger fires)
+- No llms.txt or website deliverable beyond the DR-001 one-time Lovable fix.
+- No scraping/monitoring of AI answers; no "AI visibility score."
+- No changes to the locked beta report format (tags/labels/sections).
+- No new schema, tiers, or entity graphs beyond the canonical record.
+- No monitoring implementation before a trust receipt exists.