Drift Radar

Cross-model divergence + brand-silence radar. Built on Peec AI MCP. #builtWithPeec

Drift Radar finds the prompts where ChatGPT, Gemini and AI Overview describe the same brand differently – and the prompts where they stay brand-silent entirely. Both are brand risks, and both are invisible to classic visibility metrics.

Live report: drift-radar.pages.dev

Submission for the Peec AI MCP Challenge 2026 · »Competitive Analysis« category winner.

---

What this found

Three real moments from the Pferdegold tracking set on 2026-05-23 → 2026-05-25:

0.91 divergence on »Welche Gelenkpellets für Pferde sind in Deutschland am besten bewertet?«
Gemini answers »Pferdegold Gelenke« at #1. ChatGPT answers Fohlenhof, Equine Nutritional, Pavo – Pferdegold not in the top response. AI Overview: no answer returned. Same question. Three brands at position one.

71.4 % own-brand silence
45 of 63 tracked prompts mention Pferdegold on zero of the three engines. The category is being answered – by other brands or by nobody.

17 drifting prompts at ≥ 0.30 divergence
Every one of them is a prompt where the brand is visible on one engine and invisible on another. Each gets a Wilson-95 % CI, a competitor-cited set, a content brief.

See the live report →

---

What it is

Three outputs, one daily workflow:

1. Divergence heatmap – 50 prompts × 3 engines, scored by range + coefficient of variation of own-brand visibility. Prompts with divergence ≥ 0.30 are flagged drifting.
2. Brand-silence split – own_only (competitors cited, you absent) vs full (category silent, first-mover real estate).
3. Content-brief library – one structured Markdown/PDF brief per drifting prompt, with Wilson-95 % CIs, narrative claim matrix, gap-URL excerpts and search-intent context.

How it runs

• Scheduled: .github/workflows/drift-radar-weekly.yml – Daily 02:00 UTC. Pulls fresh data via the Peec REST API (Customer API v1, x-api-key), recomputes metrics, rebuilds the site, commits. Cloudflare Pages auto-deploys.
• On-demand: gh workflow run drift-radar-weekly.yml or the »Re-run« button on the site.
• Peec-native prompt: slash_commands/peec_drift_radar.md – ready to ship as a Peec slash-command alongside /peec_engine_scorecard and /peec_campaign_tracker.

Repository layout

├── src/                       Astro site (the report UI)
├── public/                    Static assets incl. /downloads/*.csv, *.xlsx, *.pdf, *.rss
├── pipeline/
│   ├── fetch_peec.py          Peec REST API pull (skips cleanly if no API key)
│   ├── run.py                 Metrics + Wilson CIs + silence classification
│   ├── extract_claims.py      Claude Haiku narrative-claim extraction (optional)
│   ├── export_downloads.py    CSV / XLSX / PDF briefs / RSS exports
│   ├── sync_to_app.py         Copy JSON outputs into src/data/ for the build
│   ├── data/raw/              Cached Peec API responses (seed set committed)
│   └── data/ui/               drift_radar.json consumed by src/pages/index.astro
├── slash_commands/
│   └── peec_drift_radar.md    Peec-native slash-command spec
└── .github/workflows/
    └── drift-radar-weekly.yml Daily refresh + deploy (filename historic; cron is daily)

Local development

# Site
npm install
npm run dev            # localhost:4321

# Pipeline
cd pipeline
pip install -r requirements.txt
python run.py          # rebuild metrics from cached raw/
python sync_to_app.py  # copy outputs into ../src/data/

Secrets

GitHub repo → Settings → Secrets and variables → Actions:

Name	Needed for	Required
PEEC_API_KEY	Live Peec API pull (Customer API v1, available on Pro plan)	optional – without it the action falls back to the cached pipeline/data/raw/*.json
PEEC_PROJECT_ID	Scope the pull to one Peec project	optional, has a default for the Pferdegold case
ANTHROPIC_API_KEY	Narrative-claim extraction via Claude Haiku	optional

No Cloudflare secrets required – deployment uses Cloudflare Pages' GitHub integration.

Methodology

• Divergence = 0.7 × range(visibility) + 0.3 × min(CV, 2.0) / 2.0 across the active engines, per prompt.
• Wilson 95 % CI on raw visibility_count / visibility_total per prompt × model – makes N = 3 chats honest.
• Silence split from all-brands report: distinguishes own-only silence from category-wide silence so content teams can triage.
• Reference: Ethan Smith, »Demystifying Randomness in AI Search« (graphite.io).

Submission stand and post-submission updates

The Peec MCP Challenge cutoff was 2026-04-26 23:59 MESZ. The exact submission stand is tagged at submission → commit 4f4c5e1.

Updates after the cutoff (transparent log):

• 2026-04-27 – Activated the daily Peec REST pull. The pipeline was already in place at submission (see fetch_peec.py, .github/workflows/drift-radar-weekly.yml); only the API key arrived after the deadline. Switched the cron from weekly to daily (was a stub) so the rolling 3-day window matches the methodology copy in the report.
• 2026-04-27 – Two UI label fixes for consistency with the now-active daily cadence: weekly cron → daily cron in the scan strip; added snapshot from 2026-04-23 annotation to the Peec-Actions section because the REST API does not yet expose /actions, so that data stays frozen at the MCP snapshot.

Diff: 4f4c5e1...main. Demo video and submission text describe the submission stand; the live site shows the daily-refreshed state.

License

MIT. The Peec MCP Challenge IP-clause grants Peec a royalty-free worldwide licence to use the work.

---

Maintainer: Florian Meier · Challenge submission, April 2026 · #builtWithPeec