Scale & reliability

Capacity by plan

A provider check is one question asked to one AI. Monthly counts reset on the 1st. When the allowance is used up, new audits are refused. Existing audit results stay readable and exportable.

A full Pro audit is 20 questions × 10 selected providers = 200 checks. Twenty websites at that depth every day would exceed the 10,000 monthly cap. Scheduled monitoring rotates questions so a month of runs fits the allowance. The dashboard shows which questions the latest run actually asked.

How a high-volume run is scheduled

• Audits sit in a database queue. Any worker can claim the next queued row.
• One website, one active audit. A second click or a scheduled tick joins the run that is already queued or running.
• Provider calls inside a run fan out with a concurrency of 20.
• A running worker must heartbeat. Silence longer than 180 seconds is treated as a dead job.
• A queued job nobody claims for 15 minutes is cancelled so it cannot block the next start.
• Maintenance mode refuses new manual and scheduled audits until it is switched off.

Retries and rate limits

A single provider call is tried up to three times. Only temporary failures retry: HTTP 408, 409, 425, 429, 500, 502, 503, 504, and common timeout or throttle text. Permanent errors fail that provider and the scan can finish as partial instead of inventing an answer.

Backoff is 1.5 × 2^attempt seconds, or the provider's Retry-After header when it sends one, capped at 60 seconds, plus a short jitter. Failed, timed-out, cancelled, or partial audits can be retried from the dashboard. A retry replays the stored input — it does not silently pick up settings you changed later.