Graze tests one idea at a time about how a feed should be built, on real feeds whose creators lend a slice of every page to the experiment. Each test runs until readers have told us something, and every result — including the ones that found nothing — is here.
Gathering. The rate is published when the test settles, not before — a running total invites calling it early, which is the mistake this whole method is built to avoid.
Chosen because balancing recent and popular and how much likes count has looked promising across 5 completed experiment(s).
Layers readers measurably preferred. If you run a feed on Graze you can add any of these to it from your feed’s Gym page.
Readers preferred it, winning 58 per cent of the visits where they engaged with one side or the other.
Readers preferred it, winning 60 per cent of the visits where they engaged with one side or the other.
Readers preferred it, winning 58 per cent of the visits where they engaged with one side or the other.
Readers preferred it, winning 60 per cent of the visits where they engaged with one side or the other.
Readers of one feed preferred it, winning 59 per cent of the visits where they engaged with one side or the other. That feed carried 69 per cent of the evidence and the other feeds did not show the same thing, so this is a result about that audience rather than about feeds in general.
Readers of one feed preferred it, winning 61 per cent of the visits where they engaged with one side or the other. That feed carried 66 per cent of the evidence and the other feeds did not show the same thing, so this is a result about that audience rather than about feeds in general.
Carried over from the previous system. Readers preferred it 27 visits to 9, which is just short of the bar, and the instrument that would have judged it was never certified.
A creator lends a share of every page — a quarter by default — to the gym. In those slots, readers see the layer being tested alongside what the feed would otherwise have shown, mixed together so neither gets a better position than the other.
A visit is one reader, one feed, one sitting. A visit is decisive when the reader engaged with one side more than the other; most visits are not, because most people read without acting, and those count towards nothing.
A test ends when readers have preferred one side clearly, when any difference is too small to matter, or after three weeks — in which case the honest answer is that the difference is smaller than this much traffic can show, and we say so.
Nothing changes a feed automatically. A creator adds a proven layer if they want it.
On one page in ten, both halves are the same ranking. If our method were favouring a position rather than a design, that check would drift away from half and every result here would be suspect. So we watch it, in public, whether or not it is flattering.
The instrument is even: on checks where both halves are the same ranking, one half wins 51 per cent of the time over 387 visits.
Rarer actions count for more, because a signal people give seldom says more than one they give constantly. These are measured from the last 30 days across all of Graze, not chosen.