Kann ich 100 Questions vibecoden?

Build me a local AI visibility benchmark for one brand. Requirements:

- Use Node 22, TypeScript, official provider SDKs, SQLite, and a CLI.
- `benchmark --domain example.com --description "..."` creates one immutable run.
- Generate 25 buyer questions from the domain and description, or accept a JSON question file.
- Ask the exact same questions through OpenAI, Anthropic, Gemini, and xAI.
- Use each provider's supported web-search or grounding tool; keys live only in `.env`.
- Limit concurrency per provider, retry transient failures, and preserve failed cells in the report.
- Store prompts, raw answers, citations, timestamps, model ids, and errors in SQLite.
- Detect exact and case-insensitive brand mentions; allow aliases in a config file.
- Extract named competitors with one structured LLM pass after all answers are stored.
- Normalize citation URLs by hostname, canonical URL, and stripped tracking parameters.
- Compute visibility by provider, answer coverage, owned-domain citation rate, and top sources.
- Show missed questions where competitors appear but the target brand does not.
- Render a self-contained static HTML report with filters and expandable raw evidence.
- Export questions, answer metrics, competitors, and citations as CSV files.
- Every aggregate metric must link back to the answer rows used to calculate it.
- Out of scope: accounts, billing, teams, scheduled monitoring, and recommendation generation.
- Include fixture-based tests for mention detection, URL normalization, and metric calculations.
- README: setup, provider-specific grounding caveats, estimated API cost, and exact run commands.

Antworte mir auf Deutsch. Oberfläche, Kommentare und README auf Deutsch, alle Bezeichner wie Tabellen, Spalten, Variablen, Routen und Dateinamen bleiben Englisch.