The AI report
What the AI adds on top of the engine — and exactly what the model does and does not see.
What the AI adds#
The deterministic engine discovers everything; the AI explains it. On top of the raw findings you get an executive summary, the top issues with the mechanism behind each (why it slows the page, not just that it does), a per-category deep dive, and prioritized recommendations (P0–P3, each with an effort estimate). Every claim links back to the finding ids it came from.
The model is under hard instructions to never invent an issue, number, request, or domain that isn't in the engine's findings — if it isn't in the deterministic report, it can't be in the AI report.
What the model sees — and never sees#
- Sees: the engine's findings, the score breakdown, and page metadata (counts, sizes, timings, domains).
- Never sees: your raw HAR, request/response bodies, headers, cookies, or anything you typed.
The API key lives server-side behind a single gateway — it is never in the browser, and all model calls pass through per-organization rate limits, a monthly spend cap, prompt redaction, and an audit log.
The source chip & fallback#
The report header shows where the words came from: OpenAI · model when the AI wrote them, or Deterministic fallback when a limit was hit or the model was unavailable. The fallback is generated from the same findings, so you always get a complete report — the AI is an enhancement, never a dependency. See Reading a report for the deterministic layer itself.