How we check whether AI services are down

A status checker is only worth using if you know what it can and cannot tell you. This is the whole method.

Two signals, one verdict

Every five minutes a scheduled job on Cloudflare’s network gathers two kinds of evidence for each service. The first is the provider’s own status page, read by machine where the provider publishes a feed we can use. The second is our own test: we send a plain request to the service’s website and, where it has one, its API, and see whether anything answers.

The two are combined into one of three verdicts. Operational means nothing is reported and our checks get through. Having issues means the provider reports a problem with part of the service, or our checks have failed more than once in a row. Down means the provider reports a major outage of its core service, or, for services with no readable status feed, our checks have failed for about fifteen minutes in a row.

What counts as reachable

Many AI sites refuse automated requests. A reply that says “sign in” or “you look like a robot” still proves the service is answering, so we count it as reachable. Only a server error, a timeout or a failed connection counts as a failure. We also ignore a check that has never once succeeded from our side, because that points to us being blocked rather than the service being broken.

Why we are careful about calling something down

Our own check on its own can never overrule a provider’s status page that says everything is fine. It can move a service to “having issues”, never to “down”. The most common mistake in a status checker is reporting a healthy site as dead because it dislikes scripts, and this rule is there to stop it.

The uptime figure

The 30-day figure is the share of our five-minute checks in which the service was fully operational. A partial problem counts against it. It is not the same as the provider’s own uptime promise and should not be read as one. It stays hidden until we have at least three days of data, and until then the page says it is still building history.

What we cannot see

We cannot see your account, your network, your region, or whether you have reached a usage limit. We cannot tell whether answers are slow or poor, only whether the service responds. Microsoft Copilot, Grok and DeepSeek publish no status feed we can read, so those verdicts rely on our own checks and say so on the page. Google publishes no feed for the Gemini app itself, so for Gemini we read Google Cloud’s incident list for Gemini and Vertex AI and combine it with our own checks.

Data for developers

The same data is available as JSON at /api/status.json, with one service per file at /api/chatgpt.json and so on. It is free for reasonable use and may be rate limited if it is abused.