Skeletons reduce subjective wait estimates more than spinners

What it is

Clock time can be judged as two different lengths. When structure is predictable and both treatments are legal, a skeleton typically yields a shorter perceived duration than a spinner. A spinner is almost a clock: it invites people to watch “still waiting”. A skeleton fills the same interval with the forthcoming layout, so attention is spent partitioning the page and less of it is spent timing. This leaf does not decide whether a skeleton is allowed — admission is still predictable structure. It says that once the choice is open, the extra yield is pulling the wait off the timing task.

Why it happens

In prospective timing, attention acts like a gate. While the gate faces time, pacemaker pulses are counted more faithfully and the interval is judged long; while the gate is occupied by other work, pulses are missed and the interval is judged short. A spinner is a nearly pure temporal marker: a shape cycles, content slots stay empty, and it is easy to leave attention on “how many turns”. A skeleton is a filled interval, but what fills it is not stray flash — it is spatial partition work for objects that are about to appear. That work is visual processing, and it pulls the gate off timing for a stretch.

This does not automatically contradict the classic finding that filled intervals are judged longer. In the lab, when people compare empty versus filled spans as a timing task, the filled span often wins because timing is the task and the filler becomes events to count. In a UI wait the task is “is the content ready”; timing is a side job. When layout rehearsal steals that side job, the wait feels short; when a clock holds it, the wait feels long. For the skeleton to pull attention away, the fill has to look like content rather than like another clock — a full-screen unstructured shimmer puts attention back on time.

Studying it

Pair the same objective delay with a spinner or a skeleton and collect judged duration. Zakay and Block’s attentional-gate model treats “am I monitoring time” as the switch on the estimate; Thomas and Weaver’s filled-interval judgements warn that fill itself can be counted as stimulation. In HCI, Harrison and colleagues showed that progress-bar motion can push subjective speed even when clock time is fixed. Skeleton versus spinner puts “content rehearsal” against “pure time marker” inside one wait.

Independent variables: indicator type (skeleton / spinner / blank), delay length, how well the skeleton matches the final layout, whether people are told to estimate how long they waited. Dependent variables: verbal or scale duration estimates, ratio to clock time, reports of “I was watching the time”, time to find a target after arrival.

Prospective judgements (I know I am waiting and may be asked how long) and retrospective ones (only asked afterwards) bias in different directions. Product waits are almost always prospective; a lab that switches to retrospective cannot underwrite the selection. A mismatch jump on the last frame adds a separate time cost and can wipe “the wait felt short”, so fit must be a covariate, not omitted from the main effect.

Where it stops holding

When structure is unknown, buying subjective speed with a skeleton is paying in jumps; the ledger does not close. Once a wait crosses the attention-holding window, people start deciding whether to leave, and a few subjectively saved seconds do not cover “I have no idea how much remains” — progress or stages are needed, not a spinner swapped for a skeleton. Filled intervals can be judged longer in pure timing tasks; taking those filled-duration numbers as “skeletons are always faster” reverses the direction. On assistive channels a skeleton often has no equivalent; a visual advantage in judged duration is not shared with a screen reader. If the skeleton’s own pulse is noisy enough to act like an alarm, attention returns to time and the advantage dies.

Applying it

• Only when structure is predictable and both treatments are tenable, default to a skeleton — to occupy attention, not to decorate.
• Make the skeleton read as layout: header, row and image slots. Do not turn it into a fancier clock.
• Do not let this leaf cover long jobs. A lower duration estimate is not progress information; exports and installs still need a determinate bar.
• How to check: same action, same injected delay, half the people see a spinner and half a skeleton; ask immediately afterwards “about how many seconds was that”. The skeleton group’s estimates should be systematically shorter, and arrival should not jump, or the gain was not actually collected.

Related

• Same group: I2.01.1 Use a skeleton when the forthcoming structure is predictable · I2.01.2 When structure cannot be predicted, use an indeterminate indicator · I2.01.3 A skeleton that does not match real structure produces a jump · I2.01.5 An indicator on a very short load creates needless flicker · I2.01.6 A spinner carries no progress and, used for long, is read as frozen · I2.01.7 Skeleton motion needs one shared cadence; mixed breathing looks cheap
• Nearby: I2.07 Perceived performance · E6.07 Skeleton screens · I1.02 Continuity-of-thought threshold
• Search terms: perceived duration · filled interval · attentional-gate model