AI Checks

AI checks let you assert things that are hard to express as a status code or a keyword. You write what "correct" looks like in plain language, and on each run an AI model evaluates your page or API response against that description — passing or failing with a short explanation.

How it works

On each run the checker fetches your target (a rendered page or an API response), then asks an AI model to evaluate your natural-language assertion against the content. It returns a pass/fail verdict plus a rationale so you can see why it decided that.

This shines for checks that resist rigid rules:

  • "The pricing page shows three plans and none of them say Coming soon."
  • "The homepage headline is in English and mentions our product name."
  • "The checkout confirmation includes an order number and a total in USD."
  • "The API response returns a list of at least five products, each with a non-empty name and price."

Because a model call is heavier than a fetch, AI checks run on a moderate cadence (configurable). Keep assertions specific and observable so verdicts stay consistent.

What triggers an alert

  • The AI evaluates the assertion as false for the current content.
  • The target is unreachable or returns an error before evaluation.
  • Optionally, low-confidence verdicts can be surfaced as warnings for you to review.

The rationale is attached to the incident so you can confirm the call quickly.

Setting it up

  1. Add monitor and choose AI Check.
  2. Enter the target URL (page or API endpoint).
  3. Write your assertion in plain language — be specific about what must be true.
  4. Set the check interval.
  5. Attach notifications.

Phrase assertions around things that are visibly true or false on the page. Vague or subjective prompts ("does it look good?") produce inconsistent verdicts — prefer concrete, checkable statements.