CitedOS measures three engines on every paid plan, and reads all of them through their official APIs. It never scrapes the consumer ChatGPT, Gemini or Claude apps — that's brittle and against their terms. The only scraping is of the public source pages the engines cite. The free audit additionally checks Perplexity (its official Sonar API). Google AI Overviews is next: its collector is built on licensed SERP data but not yet enabled for measurement. No other engine is measured yet.
The three engines
| Engine | How CitedOS reads it | What it parses |
|---|---|---|
| ChatGPT | OpenAI web-search tool (API surface — a proxy) | Answer text + cited URLs; fan-out is simulated and validated |
| Gemini | google_search grounding tool | groundingMetadata: real sub-queries, source URIs/titles, text→source mapping |
| Claude | Claude web-search tool | url_citation annotations with start/end text spans |
Grounding metadata
For Gemini and Claude, the APIs return which part of the answer each source supports. CitedOS parses Gemini's groundingSupports (text→source mapping) and webSearchQueries (the real sub-queries the engine fanned out into), and Claude's citation spans. This is how the answer feed can show you exactly which source backed which claim.
Why no consumer-app scraping
- Scraping the consumer chatbots violates their terms of service and is technically brittle — it breaks on every UI change.
- The official APIs return clean, structured citations and grounding metadata that a scraped screen never could.
- The one thing CitedOS does scrape — via a managed scraping API — is the public web pages the engines cite (G2, blogs, Reddit, docs). Those feed Module 2.
Fan-out — the way an engine rewrites one prompt into several sub-queries — is tracked where the API exposes it. See the glossary for fan-out, grounding and related terms.