# Goose ARWP — Product Line Reviewed: **2026-09-09** after the 10-role proof-first stress test. > **Get Found.** **Goose ARWP** is the public product. **Agent-Ready Web Profile (ARWP)** is the technical foundation. Goose is an **evidence-backed Search and AI discoverability operations** product: it helps a website owner choose a small number of interventions worth trying, verify the exact deployment, preserve provider-native outcomes and decide what to keep, revise or stop. It is not another AI-visibility score, generic SEO checklist, prompt-tracking suite, protocol catalog or content-volume system. ## Public product journey A first-time user should not need to choose between Resolver, BraidGraph, Growth, Site Focus, Evidence Lab, Radar, Repository Mapper, Project Maturity Surfaces or Watch. The public journey is: ```text GOAL / DEMAND ↓ INSPECT ↓ DECIDE ONE PRIMARY MOVE ↓ CHANGE ↓ VERIFY EXACT DEPLOYMENT ↓ MEASURE PROVIDER-NATIVE OUTCOME ↓ KEEP / REVISE / STOP ``` The underlying technical sequence remains compatible with `goal -> inspect -> decide -> change -> prove -> watch`; the user-facing result is a decision cycle, not a module selector. ### Goal / demand Capture only owner context that materially changes the recommendation: audience/problem, useful action, market/language where relevant and owner-side measurement availability. Combine this with observed demand/intent/page-job evidence. Owner-declared context is not independent evidence. ### Inspect Observe the real site/repository and only the evidence needed to decide whether a candidate intervention applies: - crawl/index eligibility; - Site Focus and Intent Ownership; - page/job conflicts and demand hypotheses; - visible facts, entities and structured-data parity; - genuine datasets/Data Authority where relevant; - crawler/access policy; - Search/AI discovery state; - source ownership and deployment identity. ### Decide Return **one primary intervention** when possible, with up to three only when the actions are genuinely independent. Every action must state: - why it applies to this site; - demand/evidence basis; - source/confidence class; - exact implementation verification; - the outcome stage that would matter; - what would falsify the recommendation; - what remains unknown; - what the action does **not** prove. ### Change Use bounded, reviewable transformations only where source ownership is known. Editorial truth, legal/policy claims, authenticated provider settings and ambiguous ownership remain gated. ### Prove / measure Keep separate: - implementation verification; - production/deployment parity; - index/eligibility; - Search/AI exposure; - citations/references; - brand mentions; - visits/referrals; - useful actions/conversions. A correct implementation is not outcome evidence. ### Watch / decide The reason to return is a changed decision state: new provider evidence, a mature observation window, deployment drift, stale guidance or a review-due recommendation. Negative/neutral outcomes remain visible. --- # Canonical product primitive: the intervention chain Goose should converge on one outcome-bearing chain: ```text SITE GOAL / DEMAND ↓ OBSERVED STATE ↓ APPLICABLE RECOMMENDATION ↓ EXACT SOURCE / REPOSITORY CHANGE ↓ DEPLOYMENT PARITY ↓ PROVIDER-NATIVE OUTCOME ↓ REVIEW DECISION ``` Controlled Cohorts / Growth Experiments plus Change Receipts and outcome evidence should own runtime intervention state. Program registries may define hypotheses and candidate sites, but should reference canonical experiment IDs instead of duplicating `planned/hold/observing/reviewed` truth. Evidence Lab, Winner Observatory and the future Proof Board are views over this evidence, not parallel proof systems. --- # Internal architecture ## BraidGraph BraidGraph remains an internal lineage/index primitive connecting source/rule evidence to recommendations, source ownership, changes, verification and outcomes. Its value is traceability and reverse impact. It is not the user-facing product and does not infer causality. ## Supporting engines | Engine | Product job | | --- | --- | | Site Focus + Intent Ownership | Define the problem, demand and canonical page/job boundary. | | Growth / Recommendation Registry | Supply source-backed candidate interventions. | | Controlled Cohorts / Growth Experiments | Freeze intervention design and review outcomes. | | Evidence / Change Receipts | Preserve exact observations, changes and deployment identity. | | provider-native evidence imports | Preserve Search/AI/referral evidence without flattening semantics. | | Repository Mapper / Transformation Packs | Map rendered surfaces to exact sources and prepare bounded changes when needed. | | BraidGraph | Preserve lineage and reverse-impact relationships. | | Resolver / protocol adapters | Supporting machine-interface discovery and decision-quality work. | | Trend Intelligence | Keep upstream recommendations current. | | Project Maturity Surfaces | Project identity/rights/governance support, not discoverability proof. | | Portfolio Watch | Later multi-site review/maintenance composition. | Internal engines should not silently become first-run product choices. --- # Market boundary By 2026, established SEO/GEO products already provide prompt tracking, citations, mentions, competitors, AI visibility scores and large response/prompt datasets. Goose should not try to win on prompt corpus size or another visibility dashboard. The plausible differentiation is the intervention history: > **why this site should try this change → what exact change went live → what provider-native evidence followed → what decision was made → what must be re-reviewed when guidance changes.** The architecture is copyable. The potential moat is an accumulating, provenance-rich corpus of interventions and outcomes across sites, including null and negative results. --- # Provider-native evidence Use a funnel, not a score: `access -> index/eligibility -> exposure -> citation/reference -> brand mention -> visit -> useful action`. ### Google Search Console generative-AI reports are exposure/impression evidence with provider-supported page/country/device/time dimensions. Google does not require special AI markup for AI Overviews or AI Mode. ### Bing / Microsoft Bing AI Performance is citation/grounding evidence: citations, cited pages, grounding-query/page relationships and provider-native intent/topic/query-scoped Citation Share. It is not ranking/authority evidence and is aggregated/sampled. ### OpenAI / ChatGPT OAI-SearchBot access, public Search eligibility, bounded observed citations and referrals are separate evidence stages. GPTBot training control is separate. ### Perplexity / other answer engines Keep crawler/indexing policy and bounded observed citations/referrals where real. Do not invent owner telemetry that is not actually exposed. ## Provider-native evidence — active-cohort only Do not build a generalized provider abstraction ahead of evidence. Preserve raw owner exports and add/harden adapters when an active proof cohort has real data that cannot be represented safely today. --- # Product wedge ## P0 — First reviewed real-site proof loop The current product P0 is **not** the Proof Board. It is the first completed chain from recommendation to exact deployed change to provider-native outcome to reviewed decision. Ptichi’s frozen 12-treatment / 6-control cohort remains the canonical first attempt and stays `measurement-hold` until exact production parity is verified. No larger replacement cohort should be manufactured for appearance. ## P1 gated — Minimal proof rendering After one reviewed loop exists, #93 should render one evidence-derived proof card before any portfolio dashboard is built. The card should show site goal, intervention, evidence/demand basis, exact deployment state, provider-native outcome, unknowns/confounders and the review decision. ## P1 — Replicate the contract Apply the same proof contract to two contrasting owned sites with real owner evidence. The goal is to test generality, not fill a dashboard. ## P1 — Minimal first-run / Get Found Brief Then use #92/#94/#96 to expose URL + minimal owner context -> inspect -> one primary recommendation -> proof plan. Site Focus, Intent Ownership, Growth and evidence engines remain underneath. ## Later — Portfolio operator value After proof generalizes, the strongest paying-user hypothesis is an agency, maintainer or small team responsible for multiple sites. Potential paid value: - recurring evidence/review; - multi-site source/rule impact; - owner-data connectors; - deployment/evidence queues; - policy/governance; - reviewed remediation waves; - intervention history and learned applicability. This buyer hypothesis remains unvalidated until independent repeat use exists. --- # Stop rules Until the proof success gate is met: - do not build a rich Proof Board before the first reviewed loop; - do not add new public modules merely to make the project look mature; - do not expand Resolver/ARD/MCP/WebMCP/transact/cross-lingual breadth without a blocked real workflow; - do not treat the stratified agentic-web corpus as product P0; - do not add Trend/Radar breadth beyond maintenance; - do not add transformation-pack breadth without observed blocked target work; - do not add AI-specific files/markup as ranking tactics; - do not expand URLs before demand/intent/page-value gates; - do not use DOI/trust/policy/asset/page counts as evidence of discoverability value; - do not collapse provider evidence into a universal AI score; - do not create parallel experiment lifecycle truth; - do not hide neutral or negative outcomes. # Proof success gate Before the next major product expansion Goose needs: 1. one fully reviewed real-site intervention with exact deployment parity and provider-native outcome evidence; 2. three owned sites with canonical experiment/evidence state; 3. at least two sites with real provider-native Search/AI evidence; 4. at least one retained neutral/negative result; 5. one minimal proof artifact generated from committed evidence; 6. one demonstrated decision advantage over a generic checklist; 7. no major new capability justified by feature count or speculative maturity. If repeated well-run interventions cannot reach this gate, Goose should contract its scope rather than add product surface.