Insights

method

How these numbers are made

Every number on this page comes from one archive: a scheduled job records the repository's traffic, stars, forks, contributors and releases once a day and commits them to a public data branch. The page reads a single bundle built from that archive at deploy time. Nothing is estimated, and nothing is fetched from anywhere but this site.

Traffic

Page views and repository clones with their unique-visitor counts. A break in a line means nothing was measured at that point on the axis, a day, or a whole week once the bundle is bucketed by week, not that the figure fell to zero. GitHub reports views and clones from two separate endpoints and omits a day entirely when it recorded no traffic, so a quiet day arrives in the Reach charts as a break rather than as a zero, and the two charts can honestly disagree about which days exist. A break is not, on its own, evidence that the collector failed.

One caveat belongs to the clone line alone: GitHub counts this repository's own automation in it. Every actions/checkout, in every job of every workflow run, registers as a clone, so on a day of heavy continuous integration the line measures this project's build pipeline rather than its audience and can sit an order of magnitude above the days on either side. Clones running well above one per unique cloner is the signature. Page views are unaffected, because a checkout loads no page. The CI activity chart in Delivery is this repository's own workflow-run count, so the overlap can be read directly rather than taken on trust. Its history reaches further back than traffic collection does, because the Actions API can be asked about the past and the traffic endpoints cannot, so each chart is drawn across its own measured days rather than stretched over another chart's; hovering lands on the same day in every chart whose span covers it, and dragging sideways across one puts them all on the same span. The two are plotted separately and never combined: no fixed number of clones follows from a run, and subtracting an estimate would put a model on a page whose whole claim is that every figure on it was measured.

Cumulative counts

Stars, forks and watchers are cumulative point-in-time totals: a step the archive did not measure is a hole in the record and never a fall back to a lower count, whether that step is a day or a whole week once the bundle is bucketed by week, and wherever a line is drawn that hole is a break in it rather than a dip. A star count does not return to zero, so it is the easiest gap on the page to misread as a collapse. A day on which nobody starred is not such a gap and carries a row of its own: the total that day is known, and it is the previous day's. Star and fork history from before daily collection began is reconstructed from GitHub's permanent per-star and per-fork timestamps, which list only stars and forks that still exist, so a star since withdrawn is invisible there and those earlier totals are a lower bound rather than what the counter read at the time. The watcher count has no such record: GitHub publishes no history of it at all, so that line begins on the day the collector first ran and cannot be reconstructed any further back. The Following group is built from the rows the archive holds and nothing else: when the history is too short to make a line, a card states its reading and its reasons instead of drawing one, and how a chart's axes are scaled is stated beside the chart itself.

Referrers and paths

Referring sites and the most-visited paths in the repository, both ranked in the Reach group. GitHub reports these only as a trailing 14-day aggregate, so every figure in those two cards covers the two weeks up to its snapshot date, and there is no daily resolution to be had.

Opt-in pings

Most of this page measures this repository on GitHub. The opt-in figures in the Adoption group measure something else: the Backspace instances people actually run. A self-hosted instance can choose to send one small ping a day, and most do not, so every one of those figures is a lower bound rather than a total. An instance that never opts in is invisible, counts arrive rounded to two significant digits, and an instance is counted only once it has reported on two separate days inside the last thirty. A ping carries no domain, no instance or account name, no message content, no file names and no address. The receiver does see the connecting address, as any server does, and turns it into a two-letter country code before the ping is stored; the address itself is never written down or published, which is where the country ranking in Adoption comes from.

Nothing is charted until at least ten instances have reported inside the same seven days. Three points plotted as a line claim a shape the data does not have, so below that mark the Adoption group states the count instead of drawing it. The figures themselves are public from the first archived ping either way, in the plain tables.

Provenance and coverage

This page prints a figure in three places: the four in the row at the top, the one that heads each group, and the one on each of the smaller cards. Each of those carries a 30-day change beside it, and the change is shown only when the archive covers the window it describes. Otherwise it reads as a dash and says why, and a figure that could only be assembled from a different span is labelled with the span it actually covers.

archive history checking bundle built checking collector last ran checking last successful run checking

Prefer the numbers to the charts? The same archive is published as plain tables and as one machine-readable file, data.json. Both are built from this page's own bundle, so they cannot disagree with what is drawn here.