# Reproduce This: GEO Survey Pilot - 3 Sites - 2026-06-15

This pilot measures AI infrastructure readiness for 3 sites spanning Government, Technology, and News industries. All scoring runs through canonical edge functions (collect-comparison-receipt -> score-from-receipt) on the AnswerShare Supabase project. Receipt collection and scoring are deterministic given the same domain state; re-running against the same domain will produce updated scores reflecting current site conditions.

---

## Pipeline

### Step 1: Collect Receipt

```bash
SERVICE_KEY=<answershare-service-role-key>

curl -s -X POST https://szhmfrouktgcbkyyslvs.supabase.co/functions/v1/collect-comparison-receipt \
  -H "Authorization: Bearer $SERVICE_KEY" \
  -H "apikey: $SERVICE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"domain": "<domain>", "force_refresh": true}' \
  --max-time 180
```

Returns: `{ "receipt_id": "<uuid>", "generated_at": "<ISO>", "summary": {...} }`

### Step 2: Score from Receipt

```bash
curl -s -X POST https://szhmfrouktgcbkyyslvs.supabase.co/functions/v1/score-from-receipt \
  -H "Authorization: Bearer $SERVICE_KEY" \
  -H "apikey: $SERVICE_KEY" \
  -H "Content-Type: application/json" \
  -d '{"receipt_id": "<uuid-from-step-1>", "owned": false}' \
  --max-time 360
```

Returns: Full score object with `columns.geo`, `columns.seo`, `dashboard.*`

---

## Model Pins (at time of scoring)

| Engine | Model |
|--------|-------|
| Perplexity | sonar-pro |
| OpenAI | gpt-4o |
| Gemini | gemini-3.1-pro-preview |
| Anthropic | claude-sonnet-4-6 |
| xAI/Grok | grok-3 |

---

## 13-Signal Thresholds

### 8 Binary Signals

| Signal | Pass Condition |
|--------|---------------|
| robots_ai_bots_allowed | robots.txt does NOT block major AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, etc.) |
| llms_txt_present | /llms.txt returns HTTP 200 with markdown content (starts with #) |
| llms_full_txt_present | /llms-full.txt returns HTTP 200 with markdown content |
| sitemap_fresh | Median lastmod across sitemap pages < 60 days; null fails |
| jsonld_structured_data | 1 or more application/ld+json blocks on homepage |
| prerendered_html | Bot-fetched homepage body >= 1500 chars with no-JS rendering |
| mcp_server_live | /.well-known/mcp.json returns HTTP 200 + application/json |
| ai_content_feed | /ai-content-index.json or equivalent returns HTTP 200 |

### 5 Quantitative Signals

| Signal | Pass Condition |
|--------|---------------|
| SGR >= 0.25 | Source grounding ratio: cited claims / total claims >= 0.25 |
| RPC > 3000 | Reported page count from sitemap > 3,000 pages |
| LMR <= 30d | Median lastmod recency <= 30 days; null fails |
| RR >= 0.45 | Readability ratio (readable chars / total chars) >= 0.45 |
| RTC <= 1.0 | Response token cost (tokens * TTLB / useful chars / 4) <= 1.0 |

---

## TTFB / TTLB Methodology

- **TTFB**: `rapid_fire_ttfb_p50_ms` from 8 warm probe hits (p50 of hits 3-8 after 2 warmup hits). Threshold: 800ms. Subtract connect time for network compensation.
- **TTLB**: `rapid_fire_ttlb_p75_ms` from same probe set. Threshold: 2500ms.
- Both use the droplet egress path (143.244.215.194) for consistent measurement.

---

## GEO Bands

| Band | Score Range |
|------|------------|
| Elite | 90-100 |
| Strong | 75-89 |
| Moderate | 60-74 |
| Weak | 40-59 |
| Critical | 0-39 |

---

## How to Dispute a Score

The scoring pipeline is fully transparent. To dispute:

1. Re-run collect-comparison-receipt with `force_refresh: true` for the domain
2. Re-run score-from-receipt with the new receipt_id
3. Compare the new score against this report
4. Signal disagreements are most commonly caused by: stale sitemap data, SPA shell detection, robots.txt interpretation for AI bot blocking

---

## Known Caveats

- **Engine Audit blank**: No seeded money queries were run for this pilot. Engine-audit verdicts (AI citation visibility) require a separate 5-engine proof run.
- **AI Crawler Hits "--"**: All 3 domains are non-owned survey sites. Bot-hit telemetry is unavailable.
- **High-authority sites may score low GEO**: The 13-signal rubric measures AI infrastructure readiness, not domain authority. nasa.gov scores 6/13 despite enormous authority because it lacks llms.txt, mcp.json, fresh sitemap lastmods, and SGR signals.
- **LMR absent on many sites**: Sites with sitemaps that omit `<lastmod>` tags get LMR=null which fails per spec. bbc.com illustrates this.
- **bbc.com blocks AI crawlers**: The robots.txt explicitly blocks ClaudeBot, GPTBot, PerplexityBot, Google-Extended, anthropic-ai and many others - this directly fails the robots_ai_bots_allowed signal.
- **RTC fails all 3 sites**: The RTC metric (response token cost) is sensitive to large HTML pages with long TTLB. All 3 sites exceed the 1.0 threshold.
- **NPS legs not computed**: The muNPS/lambdaNPS/deltaNPS legs are excluded per spec. The edge function calls these legs but they return null on this project (mnps-* functions not deployed on szhmfrouktgcbkyyslvs). NPS values in the score response are informational only and not surfaced in this report.