How the internet’s own hosts are doing.

Public resolvers, content edges and the servers everything else leans on, measured continuously from boxes Helix operates and ranked on evidence you can check.

Where these figures come from

Most of the evidence comes from probe boxes Helix operates: known machines, on known networks, in known places. Latency, consistency and route stability come from those boxes alone, because they are properties of a path and a path is only meaningful between two ends you know.

Subscribers’ lines count toward one thing only: whether the host could be reached at all. Thousands of lines reaching a host from thousands of places is real evidence about its availability, so the fleet’s figure weighs at half the strength of one of our own boxes — and only once at least five different people stand behind it. Any stretch of time in which a subscriber’s own line was down is removed before anything is counted, so a host is never marked down because the people using it had a bad evening.

The order is by use; the score is quality. The top of the table is the host the most machines actually reach, whatever it scores — so the list starts with the services people depend on rather than with whichever obscure host happens to be fastest. Each row carries its quality rank beside it.

Each vantage point carries a weight, and the fold reweights over the components that actually have evidence. A target measured from too few places, or with too small a sample, is not ranked — it is listed with its figures and the reason. An absent score is never rendered as a zero.

What each column means

Used by: How many different machines reached the host on its busiest day, counted by the host census — which is what the list is ordered by. “Always” marks a host measured whatever the count says: a public resolver, or one of the sites every agent’s Internet Metrics board measures. A dash marks any other host the census has not counted above its floor in the last month: it is measured because subscribers’ own agents measure it, or because the census counted it before.
Score: The overall, out of 100, from the five columns that follow. Beside it is the host’s place when the list is ordered by score, out of every host that has a score.
Availability: Echoes delivered, of those sent. Below 300 probes it is not shown as a rate at all — the counts travel instead, because a percentage from a small sample is a number nobody can defend.
Latency: The median round trip, judged against the distance actually travelled. 140 ms across a city is broken; 140 ms across an ocean is roughly the speed of light through glass. A ranking on raw milliseconds would only rank hosts by how close they happen to be to our probes.
Consistency: Jitter, and how far the slow tail runs above the median. This is what a video call feels when the average looks fine.
Route stability: How often the route to the host changed. A path that keeps moving is a worse path than one that does not, even when both are fast.
Reach: How much of the fleet, by weight, reached the host at all.
Probes: How many echoes Helix’s own boxes sent the host over the days covered. Every figure in the row rests on them.
Vantage points: How many of Helix’s own boxes measured the host. One box’s view is one network’s view, which is why a host seen from too few is listed but not ranked.

What this is not

Not a ranking of ISPs. It ranks hosts and the paths to them. How an ISP serves the people on it is a different question with its own page and its own publication floor.
Not a status page. It folds completed days. A host that broke an hour ago is not in it, and the table prints the period it covers.
Not a synthetic benchmark. Nothing is modelled, simulated or filled in. A gap in the data is a gap on the page.