Written out in enough detail that anybody with a stopwatch can reproduce a figure, disagree with our conclusion and still use our numbers. The procedure has not changed since the first review and will be versioned here if it ever does.
Every figure on this site was recorded on the same desktop, in the same room, on the same drive. Nothing is scored before twenty hours of play, and no game is timed on a build older than the one on sale.One mid-range desktop of the kind most readers actually own rather than the kind review guides are written on: a six-core processor several years old, thirty-two gigabytes of memory, a mid-tier graphics card and a single NVMe drive with at least forty per cent free space. The operating system is reinstalled twice a year and carries nothing but the games under test and a stopwatch.
The absolute numbers therefore do not predict your machine. What they do is rank these games against each other, which is the only comparison a single tester can make honestly.
A cold run means the machine has been restarted, no other application has been opened, and the game has not been launched since the restart. A warm run is the same measurement taken immediately after quitting the game, when the operating system still holds much of it in memory. Three of each, and we publish the median.
The gap between the two is often the most revealing figure we take. A game with an eleven-second warm boot and a ninety-second cold boot is a game most of its own developers have never experienced the way a player does on a weekday evening.
Start: the double-click. Stop: the first frame that accepts and acts on input. Not the first frame drawn, not the appearance of a menu, and not the studio logo. A game that shows you a main menu you cannot yet click has not finished loading, and several in this file behave exactly that way.
Unskippable logos and splash sequences are included in the boot figure, because they are unskippable. Where a game allows them to be skipped we take the figure with the skip, and note that the option exists.
Four categories, each timed forty times and reported as a median: fast travel or area transition, respawn after death, the wait between selecting continue and regaining control, and any menu that stalls before becoming interactive. Where a game hides loading inside an animation, we time the whole animation and separately estimate the real load by watching drive activity, and we publish both.
We also record whether input is discarded during a transition, which is a distinct failure from the wait itself and appears in several reviews as a half-second of vanished clicks.
Small waits are aggregated by counting occurrences over a complete playthrough with a physical tally, which is tedious and is the reason a review takes weeks. A three-second transition repeated four hundred times is a different fact from a three-second transition, and the aggregate is where the interesting arguments live.
These totals are specific to how we played. Somebody who fast travels less, or dies less, will pay less, and the reviews say so. The count is published alongside the total precisely so a reader can substitute their own habits.
Recorded separately from the duration. A ninety-second wait with a percentage that advances is tolerable; the same wait with a spinning shape and no information produces the reasonable belief that something has broken. Two games in this file have taught players to force-quit a working process, and that is a design failure independent of speed.
A review score for the whole game: design, writing, structure, options and craft. Timings are recorded facts printed beside it, not a formula that produces it. A slow game can score well and does, twice in the current file, and where that happens the review says so in its opening paragraph rather than burying the qualification.
All timings are taken on an unmodified installation at the game's own default graphical settings. Community patches frequently improve load times substantially, and that does not excuse a shipped default. Where a widely used fix exists we may mention it in the text; no figure is adjusted for it.
Load times change under patching more than almost any other measurable property. When a patch materially alters a figure we re-time it and annotate the review with the date and both values, leaving the original visible. The shipped state was real and the record should show it.
One machine, one drive, one tester, one set of habits. Absolute figures will not match yours. Aggregates depend on how we played, which for an open world is a substantial assumption. Twenty hours is enough to learn a game's rhythm but not always enough to encounter its worst transition, and in two cases readers found slower ones than we did.
Our selection of games is also unsystematic and biased toward titles whose waiting is interesting, which should be remembered whenever we quote a median for the file as a whole.
Every game timed here was bought at retail. No developer, publisher or platform supplied a copy, funded a piece or saw anything before publication. There is no advertising on this site, no sponsorship, no affiliate arrangement and no storefront link on any page.
Figures are corrected in place with a dated note whenever a reader demonstrates an error or a patch changes the behaviour. There is deliberately no address, form or submission route anywhere on this site; corrections reach us because readers repeat measurements publicly, which has worked better than a contact page ever did.
