How AIDownCheck works

Most outage trackers answer one question: is it down? AIDownCheck answers four more: why, who else is affected, what can I use instead, and how sure are we.

Some trackers are just a tally of complaints. A report count tells you people are upset; it can't tell you what actually broke. We read the vendor's own status page and run our own checks andmap what each tool depends on — so we can tell you what's down, how we know, and what it takes down with it.

Three independent signals

For each service we combine sources that fail in different ways:

• Official status pages. We read the vendor's own status feed — via Atlassian Statuspage, Better Stack, Instatus or incident.io, and for the odd vendor that rolls its own, we read that too. It's authoritative but often lags the first minutes of an incident. Where a vendor publishes per-component status, we surface which part (API vs web app vs image) is affected.
• Independent checks.We run our own reachability and response-time check from our servers every minute. It catches problems before a status page updates — and it lets us flag when a service is reachable but slower than its own normal (“brown-outs”), which status pages never show. A single vantage point can be noisy, so we treat it conservatively.
• Your reports. First-party “is it down for you?” votes add a human signal. They're privacy-safe — we never store your IP — and used only as an early warning; a spike alone never produces a hard “down.”

The confidence engine

We never claim certainty we don't have. Every verdict carries a confidence level and shows the evidence behind it:

• High — both sources agree (e.g. the vendor confirmed an incident, or both say healthy).
• Medium— sources conflict, like an official page still showing green while our checks fail. That's the early-warning window.
• Low / Unconfirmed — we only have one usable signal, so we say so plainly.

Dependency intelligence

Modern AI tools are built on a handful of model providers. Cursor and Bolt run on Anthropic's Claude; many apps run on OpenAI. When a provider has an incident, the products built on it often degrade too — frequently before their own status pages admit anything. We map those public relationships so you can see the blast radius, always framed as likely impact, never confirmed causation. It also works in reverse: when several independent products built on the same provider degrade at once while that provider's status page is still green, that's an early signal of an upstream incident — which we record conservatively as the leading edge of detection.

Honest by default

We don't invent uptime percentages or outage durations. Historical statistics appear only once our database has accumulated real measurements — and until then we tell you exactly what we do and don't know. AIDownCheck is independent and not affiliated with any AI provider.