Profound

Build me a local AI answer engine visibility tracker for one brand. Requirements: - Node 22, TypeScript, SQLite via better-sqlite3, a CLI, and a plain server-rendered dashboard. Local only, no accounts, no telemetry. - brand.json holds my brand name, aliases, domain, and competitor names. prompts.json holds up to 40 buyer questions. - `track run` sends every prompt through OpenAI, Anthropic, Gemini, and Perplexity with each provider's web search or grounding tool enabled. Keys live in .env. - Store one immutable row per run, prompt, and provider: raw answer, cited URLs, model id, latency, and error text. Never overwrite an existing run. - Cap concurrency at 3 per provider, retry twice on 429 and 5xx with backoff, and keep failed cells visible in the report instead of dropping them. - Detect brand and competitor mentions case-insensitively using the alias list, and record the first-mention character offset as a crude prominence proxy. - Score the sentiment of each brand mention in one cheap structured pass over stored answers, after the run, never inline. - Normalize citations to hostname plus canonical path, strip tracking parameters, then compute owned-domain citation share and a top 25 sources table. - `track serve` renders visibility per provider over time, share of voice against each competitor, the sources table, and the prompts where competitors appear and I do not. - `track crawlers --log access.log` parses server logs for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended and friends, plus referral hits from chatgpt.com and perplexity.ai, and reports which URLs they touched. - Keep the bot user agent list in an editable JSON file. Log lines matching nothing get counted as unknown agents, not silently discarded. - `track export` writes runs, mentions, and citations to CSV. - Fixture tests for mention detection, URL normalization, and log parsing. - Out of scope: real consumer surface answers, prompt volume estimates, content generation agents, teams, and hosted scheduling. Do not scrape the consumer web UIs. - README: setup, per-run cost estimate, a cron line for daily runs, and a plain note that API answers only approximate what users actually see.