The Psychology of Waiting: The Two Rules Everyone Quotes Are Not Findings

Interfaces make you wait, and how they behave during the wait changes how long the wait feels. That much is real and measured.

Almost everything else you have been told about it is not. The two rules quoted most often are not experimental findings; the mechanism cited most often has never been independently replicated; and all six of the famous latency statistics trace to a vendor slide, a commissioned opinion survey, or a white paper that admits it does not know where its own number came from.

This article is mostly about that gap, because the gap is larger than the evidence.

The 0.1 / 1 / 10 second limits are a design table from 1983

You have seen them: 0.1 seconds feels instantaneous, 1 second keeps your train of thought, 10 seconds is the limit of attention. They are usually attributed to Jakob Nielsen, and Nielsen does not claim them as his own research. His 1993 text cites two sources.

Follow them back.

Miller (1968), in the AFIPS Fall Joint Computer Conference proceedings, is the origin. Miller explicitly disclaims empirical status. The thresholds are, in his words, “the best calculated guesses by the author, a behavioral scientist, who has specialized in task behavior,” and he says they “should, indeed, be verified by extended systems studies.” His own numbers are not the famous ones either: 0.1 to 0.2 seconds for keypress feedback, two seconds for routine acknowledgement, one second for page turning, and fifteen seconds as the point that rules out conversational interaction. Miller never states a 10-second limit.

Card, Robertson and Mackinlay (1991), at CHI, is the second source. The constants do appear there, in a table headed “Human Time Constants for Tuning Cognitive Co-Processor”: perceptual processing 0.1 s, immediate response 1 s, unit task 10 s. The references given are two textbooks - Card, Moran and Newell’s The Psychology of Human-Computer Interaction (1983) and Newell’s Unified Theories of Cognition (1990).

That paper contains no user experiment. We checked the body text: zero occurrences of “subject,” zero of “participant,” zero of “significan,” zero of “ANOVA,” zero of “p <.” It is a systems paper by Xerox PARC describing their own product architecture, with a design-tuning table attached.

And the ten seconds is a compression. In the 1983 source the unit task is given as about ten seconds “(say, 5 to 30 seconds).” A range of 5 to 30 became a hard limit of 10.

So: a behavioural scientist’s self-described guesses, passed through a textbook and a systems paper, hardened by repetition into a law.

The 2-second rule is a hypothesis borrowed from someone else

The other widely quoted figure is that users tolerate about two seconds. This traces to Nah’s tolerablewaiting-time study.

In it, the two seconds is hypothesis H3, derived from a 1984 recommendation by Shneiderman - not a measurement the study set out to make. The paper states: “based on Shneiderman’s recommendation, it is hypothesized that: H3: The threshold for tolerable waiting time is 2 seconds,” and later that “our results seem to support Hypothesis 3.”

The empirical support is a mode. Seventy undergraduates clicked ten hyperlinks of which only seven worked; tolerable waiting time was operationalised as the elapsed time between clicking a link and clicking STOP. For the second and third broken links, “The mode for both groups was in the interval of 2-3 seconds, with a frequency of 13 (36%) for the control group.”

A modal bin of 2 to 3 seconds, containing 36% of one 36-person group, waiting for deliberately broken links, testing a hypothesis borrowed from a 1984 book. That is the two-second rule.

What the study did establish, and it is the best finding here

The same study produced something genuinely useful. It compared groups with and without a feedback bar:

Broken link	No feedback	Feedback bar	p
First	13 s	37.6 s	< .000
Second	4 s	17 s	< .002
Third	3.3 s	6.7 s	< .004

“The inclusion of a feedback bar significantly prolonged the waiting time of users.” Showing something roughly tripled how long people would wait.

The same direction appears in the classic HP Labs work by Bouch, Kuchinsky and Bhatti (2000), with 30 participants on a shopping task. Rating pages that loaded all at once, participants called 0 to 5 seconds “high” quality and anything over 11 seconds “low.” Rating pages that loaded incrementally - banner first, then text, then graphics - the same participants called 0 to 39 seconds “high” and the low threshold moved past 56 seconds. In the authors’ words, the upper bounds “are almost 6 times higher in each case.”

Showing progressive content multiplied acceptable latency roughly sixfold. That is the strongest practical result in this field.

Two caveats belong with it. The sample was 30 men. And within the same study, how participants rated latency showed no correlation with how they behaved when given the option to bypass a delay. Stated preference and revealed behaviour came apart in the same thirty people.

Progress bars can bend perceived time, at a specific size

Harrison, Yeo and Hudson (2010), at CHI, is the source of the widely repeated claim that the right animation makes a wait feel about 11% shorter. The claim is real. It needs three corrections.

It is not pulses. It is ribbing - specifically ribbing that moves backwards and decelerates. Pulsating bars were a different study in the same paper and were not the design measured for magnitude.

The measurement runs the other way. What was measured is that a 5-second plain bar felt equivalent to a 5.61-second ribbed bar (SE = 0.14), 12.2% longer, and a 15-second plain bar felt equivalent to 16.75 seconds (SE = 0.45), 11.7% longer. Both at p < 0.001. The paper reframes this as an 11% reduction in perceived duration; the safe sentence is that a wait felt equivalent to a plain bar roughly 11 to 12% longer.

The sample for that measurement is 16 people. And an earlier study in the same paper found that at 5 seconds only a single comparison even approached significance (p < 0.1); the effect essentially lives at 15 seconds.

An earlier paper by Harrison and colleagues (2007) at UIST, with 22 participants from two computer research labs, found that accelerating progress functions were preferred over linear ones - with “no preference” the modal response at 36% of 990 paired comparisons.

Independent support comes from Ziat and colleagues (2022) in Scientific Reports, whose four within-subjects experiments held actual speed and distance constant and found progress-bar step structure altering duration estimates. Their samples are 20, 11, 21 and 16 undergraduates, with no sample-size justification and no limitations section, and their Experiment 4 found no significant step effect on throbber duration estimates at 10 to 14 seconds.

Progress indicators themselves: one clear result, one failed replication

Myers (1985), at CHI, tested 48 subjects - mostly computer science graduate students, unpaid volunteers - on tasks with and without a percent-done indicator.

Progress indicators were strongly preferred: F = 13.70, p = 0.0006. “86.1% percent of the subjects liked progress indicators,” with a mean rating of 2.94 where 1 is “Very Useful” and 9 is “Useless, Annoying.”

But the same paper is a failed replication, and says so. It set out to reproduce an earlier finding that people prefer constant to variable response time. The variability main effect was F = 1.66, p = 0.2046 - null. The hypothesised interaction was null too, F = 2.39, p = 0.1296. In the paper’s own words, “no significant preference for constant response time was shown, contrary to previously published results,” and it “failed to duplicate the results of earlier experiments.”

Skeleton screens: no evidence either way

Skeleton screens - grey placeholder shapes that mimic the page about to load - are currently the most fashionable pattern in this area. The peer-reviewed evidence for them is one four-page paper.

Mejtoft, Långström and Söderström (2018), at ECCE, compared skeleton screens against spinners with a 2.75-second artificial delay. Fourteen participants, seven per condition.

All three tests were null: time to click a link p = 0.69 (spinners 13.60 s, skeleton 15.01 s - spinner users numerically faster), perceived speed p = 0.16, ease of navigation p = 0.31. The authors’ own conclusion: “A lot more data is needed to verify the claim.”

Seven per condition is not evidence against skeleton screens. It is no evidence at all, and it is the entire peer-reviewed base for the pattern.

More recent work has moved on. Dufresne, Gutwin and Graham (2025), in Proceedings of the ACM on Human-Computer Interaction, compared “false front” placeholder pages against spinners and skeleton screens with 72 and 96 participants. False fronts produced faster completion - loading style F(2,69) = 5.05, p = 9.02e-03, η2 = 0.07, with false fronts at 1183 ms against placeholders at 1584 ms and spinners at 1651 ms. But the perceived-speed benefit appeared only at delays of 1000 ms and above; at 500 ms there was nothing.

That bound recurs. Arapakis, Bai and Cambazoglu (2014), at SIGIR, found delays under 500 ms were not easily noticeable, while delays above 1000 ms were noticed with very high likelihood. Below half a second, there is not much for loading design to do.

The labor illusion, and the replication that never came

The most quotable idea in this area is the labor illusion: showing people the work being done makes a slower service feel more valuable than an instant one.

Buell and Norton (2011), in Management Science, ran five experiments in travel search and online dating. Experiment 1, N = 266: transparency raised perceived value, F(1,212) = 10.68, p < 0.01, transparent M = 5.36 against blind M = 4.96. Experiment 2, N = 118: when forced to choose between a transparent 60-second wait and an instant competitor, 63% chose to wait, against 23% for a blind 60-second wait.

Two things must be said with it.

Its own Experiment 5 found the effect reverses. With N = 280, when outcomes were unfavourable, transparency reduced perceived value relative to instant results: at 15 seconds t(62) = 2.40, p < 0.05; at 30 seconds t(58) = 2.97, p < 0.01. “Show your work and people will wait happily” is not what the paper found.

In fifteen years it has never been directly replicated by an independent team. We searched its citation network specifically for replications and found none - the closest results are extensions, several by the original author. Its theoretical parent, the effort heuristic, was subjected to a replication attempt published in Collabra: Psychology in 2023, which found mixed results: support for one original experiment, failure on another, and in a second attempt support on liking and quality ratings but not on monetary value - which is precisely the outcome the labor illusion is usually invoked to explain.

The six statistics you should refuse

Every famous number about latency and abandonment traces to vendor material. Here is where each actually comes from.

“100 ms costs Amazon 1% of sales.” One bullet on one slide of a 2006 Stanford guest lecture by a former Amazon employee, reading “+100 ms -1% sales @ Amazon.” No sample size, no date, no scope, no confidence interval, no definition of sales. Amazon has never published the experiment.

“A 500 ms delay cost Google 20% of traffic.” The same slide, from a 2006 conference talk. What was actually done: Google increased results per page from 10 to 30, which also raised load time from 0.4 to 0.9 seconds. The manipulation changed two things at once. The 20% drop cannot be attributed to latency. It is a results-count experiment.

“47% of consumers expect a page in 2 seconds” and “40% abandon after 3 seconds.” A 2009 Forrester Consulting survey commissioned by Akamai - 1,048 online shoppers asked how long they were willing to wait, with pre-set answer options. Stated intention, not behaviour, with anchored response ranges. No website analytics were used.

“53% of mobile visits are abandoned after 3 seconds.” A 2016 DoubleClick by Google report. Its own source note reads: “Aggregated, anonymized Google Analytics data from a sample of mWeb sites opted into sharing benchmark data, n=3.7K.” The 3.7K is a count of websites, not users - it is routinely miscited as a sample of people. Self-selected, observational, no confidence interval, and subject to a known artefact: a visitor who leaves before the page finishes loading may never fire the completion beacon, which mechanically inflates the association.

The “8-second rule.” A 1999 Zona Research white paper, which does not claim to have discovered it and cannot say where it came from: “Though few can remember exactly when or where the term first gained credence, it is now widely believed…” Its behavioural data come from a 1998 holiday shopping survey of 208 respondents, and it concedes that “Due to non-standardized data… certain assumptions had to be made.” It describes 1999 dial-up conditions.

If you need a real latency-and-behaviour result, use Arapakis and colleagues: peer-reviewed, with an industrial-scale query log, holding content constant.

What this leaves you with

Show something rather than nothing. This is the best-supported claim in the field, from two directions: a feedback bar roughly tripled tolerated waiting, and incremental loading multiplied acceptable latency about sixfold.

Tell people how long it will take. Field experiments on telephone waits found expected-duration information significantly reduced how long people thought they had waited - though the published version of that work notes it also increased the negative effect of perceived waiting on evaluations, making it a double-edged tool rather than a free win.

Below about 500 ms, stop optimising the animation. Two independent studies put the noticing threshold there.

Do not quote the 0.1/1/10 limits or the 8-second rule as findings. They are, respectively, a design-tuning table citing two textbooks, and a white paper that admits it does not know where its own rule came from.

Skeleton screens are unevidenced, not disproven. Seven participants per condition tells you nothing.

For the design pattern that removed the natural stopping point, see infinite scroll. For friction deliberately introduced on the way out, see the roach motel dark pattern. For the wider pattern, see the persuasive design guide.

Limitations of this article

Two sources were read in versions other than the version of record. The tolerable-waiting-time figures come from the open conference version of that study rather than the later journal article, and the telephone-waiting statistics come from a working-paper version. The published abstract of the

latter frames expected-wait information as a double-edged effect, and we have used the published framing.

The classic samples are small and unrepresentative: 16, 20, 22, 30 men, 48 mostly computer science graduate students, 70 undergraduates, and 7 per condition.

The largest-scale latency evidence is proprietary. The SIGIR study was conducted by Yahoo Labs employees on a 30-million-query internal log that cannot be independently audited.

The labor illusion’s citation network is substantially self-referential, with most extensions authored by its originator.

We did not verify the funding statement for the 2025 false-fronts study, and we read one 2025 progress-animation study in a publisher summary rather than its PDF; its figures are not printed here. See the withheld-figures record.

No correction, retraction or expression of concern was found on any paper cited here; we checked Crossref update fields and retraction status for every DOI.