Back to Blog

You have seen the claim: human attention has collapsed to a handful of seconds, less than a goldfish. It is the most repeated statistic in web design, and there is no study behind it. Not a weak study, not a misread study. None.

That is the first of five. Before the rest, the standard we are holding things to, because a debunk that ignores its own sourcing is just a different set of confident claims. Very little research on web reading is peer reviewed. Most of the good work is eyetracking by UX consultancies and most of the rest is analytics from vendors who sell measurement. That does not make it worthless, but a consultancy heatmap cannot support a law of human behaviour, and we say which kind of source each claim rests on as we go. This is also the fifth post in our evidence series and the first where we found our own site guilty, which is at the end.

The attention-span statistic has no study behind it

The figure is usually credited to a 2015 Microsoft Canada consumer-insights report, a real document based on a survey of 2,000 Canadians, 112 of whom also had their attention measured with portable EEG.

It is not Microsoft's finding. The report carried it with a footnote to a third party, Statistic Brain, a statistics aggregator that was offline as of September 2026. Statistic Brain listed its own sources as the National Center for Biotechnology Information at the US National Library of Medicine, and the Associated Press.

In 2017 the BBC World Service programme More or Less chased it down. Simon Maybin contacted both listed sources. In his words, "neither can find any record of research that backs up the stats." Attempts to reach Statistic Brain came to nothing, and the attention researchers he spoke to had no idea where the numbers came from either.

That is worth separating from an ordinary bad statistic. An unsourced number is usually a real result whose citation got lost. Here there is no result. A figure was published by an aggregator, picked up in a Microsoft-branded PDF, and carried from there into Time, the Telegraph, the Guardian, USA Today and a shelf of business books.

The goldfish half is better still, because the research points the other way. Professor Felicity Huntingford, who by 2017 had studied fish behaviour for close to fifty years, told the BBC that goldfish "can perform all the kinds of learning that have been described for mammals and birds," and that they have become a model system for studying memory formation precisely because they learn and remember well. Her own summary was that it is an interesting irony.

Two caveats we will not skip. Microsoft never retracted anything; the report simply came off its site, so there is no tidy admission to cite. And the BBC piece is journalism, not a peer-reviewed finding. The honest framing is that the best available trace of this number ends in two institutions unable to find it, not that science has disproved it.

There is real research on attention at work. Professor Gloria Mark at UC Irvine has logged how long people stay on one screen before switching: roughly two and a half minutes in 2004, about 75 seconds by 2012, and around 47 seconds in more recent measurements. That is genuine academic work, repeatedly measured. It is also a record of task switching among people juggling jobs, not a measure of how long anyone can attend to something, and it says nothing about how long someone will read a page they actually want to read. Swapping one number in for the other is the same mistake wearing a better suit.

The reading percentage was calculated, not measured

The most quoted figure in web copywriting is that people read only about 20% to 28% of the words on a page. It is normally cited to a peer-reviewed paper in ACM Transactions on the Web by Weinreich, Obendorf, Herder and Mayer, published in 2008 and built on browsing data from 25 volunteers the authors recruited themselves, collected on desktop in the winter of 2004 and 2005.

That paper is real and its dataset is genuine. Neither figure appears in it.

Both come from a separate re-analysis Jakob Nielsen published on his own site the same year. He took the time people spent on pages and asked what share of a 593-word page someone could get through at an assumed 250 words per minute if they spent every second of the visit reading. That produced 28%. He is explicit that it is hypothetical, writing that "the total time spent on a page is definitely the upper limit of possible reading time," and noting "this wasn't an eyetracking study, so we don't know precisely how users allocated their time on the Web pages."

So 28% is a ceiling: the most someone could possibly have read, not what anyone did read. Even the 593-word page is his own figure, computed from the raw records, and it appears nowhere in the published paper, which reports 551 words with outliers trimmed or 648 without.

The 20% figure, the one the marketing industry quotes more often, has no derivation shown at all. Its entire basis is one sentence: "More realistically, users will read about 20% of the text on the average page." In fairness, run the same calculation at Nielsen's usual 200 words per minute rather than the 250 he raised it to here, and you land near 22%, so 20% is not plucked from nowhere. It is still a second hypothetical rather than a measurement, and no working is shown for it.

The chain runs: real browsing data from 25 people in 2004, then a hypothetical ceiling built on assumptions, then a downward adjustment with no shown work, then eighteen years of citation back to a peer-reviewed paper that contains neither number.

We are not saying people read every word. They plainly do not. We are saying that if you have been quoting this to a client as a measured fact from a journal, it is neither measured nor from that journal.

The F-pattern and the fold say the opposite of what they are used for

These two shape more layouts than anything else on this list, and both are routinely used to justify the opposite of what their own sources recommend.

The F-pattern is a symptom, not a target

Nielsen Norman Group's 2006 eyetracking work, with 232 users across thousands of pages, found people often scan text-heavy pages in a rough F. That is consultancy research, not peer reviewed, and the article itself calls the pattern "a rough, general shape rather than a uniform, pixel-perfect behavior."

What got lost is what NN/g themselves concluded. Their 2017 follow-up by Kara Pernice argues the F is a problem to fix rather than a shape to design toward. It is "the default pattern when there are no strong cues to attract the eyes towards meaningful information," and it appears when "your pages have big chunks of unformatted text." Their recommendation is to break it with headings, subheadings, bold, bullets and front-loaded opening sentences.

NN/g's later taxonomy names several other scanning patterns and ranks the layer-cake, where someone reads headings and subheadings and skips the text between, as the most effective. That argues for a strong heading hierarchy, not an F-shaped layout. A 2007 study by Shrestha and colleagues narrowed it further, finding that scanning depends on task and content, with image-based pages drawing attention to images rather than tracing an F. And a larger commercial analysis of 99 websites and 157,498 fixations reported no F-shaped bias at all, finding attention skewed toward the middle and slightly left. That last one is vendor research measuring arrival scanning rather than prose reading, so it narrows the claim rather than settling it.

The myth says: put everything important along the top bar and the left edge, because that is where the eyes go.

The evidence says: if your readers are tracing an F, your formatting has failed, and the fix is structure that gives their eyes somewhere better to land.

People do scroll, and have since 1997

"Nobody scrolls, put everything above the fold" was a genuine finding once, on the web of the mid-nineties. Nielsen withdrew the advice in 1997, reporting that users had started scrolling. The sources quoted in support of it today say the opposite: NN/g's 2018 scroll study states that "people will scroll if they have a reason to do it," and their 2015 piece on the fold says "users scroll when there is reason to."

What survives is subtler and more useful. That 2018 eyetracking, with 120 participants and over 130,000 fixations on desktop at 1920x1080, found about 57% of page-viewing time was spent above the fold, roughly 74% within the first two screenfuls and 81% within three. More than 42% of viewing time fell in the top fifth of the page. NN/g's own stated limitation is that the analysis disregards page length, so the result may partly reflect how long the sampled pages were.

Set against that is a single Chartbeat dataset, surfaced in Slate in 2013 and restated in Time in 2014 by Chartbeat's own chief executive, putting roughly two-thirds of time on a page below the initial screen. That is one vendor number appearing twice, not two independent findings, and it comes from a company selling attention metrics.

The likeliest reconciliation is page length: on a short page most attention lands above the fold because most of the page is above the fold, and on a long article the fold sits near the top of the body so most reading time necessarily falls below it. Neither source publishes the length distribution of its sample, so treat that as a plausible explanation rather than a demonstrated one.

The myth says: cram the important things above the fold.

The evidence says: put the highest-priority thing at the top and let the page run as long as it needs, because cramming produces exactly the dense, unformatted block that sends people back into an F.

Banner blindness is real, and broader than the version people quote

This is the one case where getting the detail right improves the advice.

The term was coined by Jan Panero Benway and David M. Lane at Rice University in 1998, in a technical-group newsletter, with a companion conference paper the same year. The newsletter is what everyone cites, and it is routinely miscredited as a peer-reviewed journal article, which it is not. The finding everyone quotes, that users located banner-style links about 58% of the time against 94% for ordinary menu links, comes from a pilot with six users. Their larger follow-up with 72 undergraduates found no significant speed benefit from a helpful banner at all.

The important part is what the original authors meant. They defined "banner" broadly rather than as advertising, and concluded the effect "is wider than just an advertising effect." They reproduced it with plain blue text that looked nothing like an ad. The peer-reviewed replication by Burke, Hornof, Nilsen and Gorman (ACM Transactions on Computer-Human Interaction, 2005, 24 eye-tracked participants) found blank grey placeholder banners were fixated about as often as real commercial ones, which is hard to square with an ad-recognition story.

Two honest complications

The mechanism is contested. Hervet, Guerard, Tremblay and Chtourou (Applied Cognitive Psychology, 2011) used eyetracking and found most participants did fixate the ads at least once, so the defensible claim is that people look without encoding or acting, rather than never looking.

And ignoring is not free. In that 2005 replication, static banners slowed visual search by 6.3% and animated banners by 7.5% against a no-banner baseline, while a target sitting next to flashing banners took about 75% longer to find than the same target without them. Those are different comparisons, which is why the figures sit an order of magnitude apart. Participants also rated their workload higher.

So the effect is not "people ignore things that look like ads." It is closer to "people ignore things sitting outside the structure they are searching," which applies to your own content too. NN/g's separate 2018 banner-blindness eyetracking (26 participants, consultancy research) reports the same for right-hand rails and other ad-typical positions. Treat that as practitioner evidence about placement, not a proven law.

The myth says: avoid anything that looks like an ad.

The evidence says: keep the emphasized item inside the link menu, which is Benway and Lane's own recommendation, and never let a primary action exist only inside a banner or a floating box. Carousels are a different argument resting on weaker evidence, so we are not basing any carousel advice on this research.

What we are not saying

The behaviour is real, even where the numbers are not. People scan rather than read, and visually isolated elements do get missed. Those hold up across every source here, weak and strong. What does not follow is any particular percentage, or the layout rules people build on top of them.

Attention concentrating near the top is the shakiest of the survivors. Its support is one consultancy study whose own limitation we quoted, and this post then argues that on long pages most reading time falls below the fold. Read it as a tendency on desktop at the page lengths studied, not a law.

Our own sources are mostly not peer reviewed. The goldfish trace is a radio programme. The fold and F-pattern numbers come from a UX consultancy. The scroll-depth counter-evidence comes from an analytics vendor whose chief executive wrote the article. They are the best evidence available on these questions, and we would rather name what they are than dress them up.

Not peer reviewed does not mean wrong. NN/g's eyetracking is careful practitioner work, and on several of these questions it is the only substantial evidence anyone has gathered. The problem is the leap from a heatmap to a law.

This evidence is old, and all of it is desktop. The primary datasets run from 1998 to 2018, the most recent eyetracked at 1920x1080. None of it describes reading on a phone in 2026. We cannot criticize other people for stale data and then exempt ourselves.

No meta-analysis exists for any of these claims. The evidence base is an accumulation of small studies, several with fewer than thirty participants. Absence of a good number is not proof of the opposite, and we are not claiming attention is provably unchanged, only that nobody has shown it changed.

What to say instead

You still need something to tell a stakeholder on Monday. None of the following are measured findings. They are what is left once the unsupported numbers are removed, and they fail safe: if the underlying research is weaker than it looks, following them still costs you nothing.

Instead of an attention budget in seconds: people arrive with a question and leave when a page stops answering it. That is why the top of the page matters, and it needs no stopwatch figure nobody can source.

Instead of a reading percentage: write as though every heading will be read and the body beneath it may not. NN/g's consultancy eyetracking suggests headings carry disproportionate weight, which is a reasonable working assumption rather than a measured law, and it drives the same discipline without a manufactured number.

Instead of designing for the F: write headings that carry meaning alone, then read only your headings and see whether the page still makes sense.

Instead of cramming above the fold: lead with the single highest-priority element and let the length follow the argument.

The unsourced statistic we published ourselves

While writing this we searched every post on this site for claims that put a number of seconds on how long a visitor will look at a page, and for unsourced reading percentages. We found two, both ours. (A separate post advises hooking a video ad in its first few seconds, which is a production convention about video retention rather than a claim about how long anyone reads, so we left it alone.)

A post about marketing Calgary gyms told readers that prospects "make a visceral judgment in 8 seconds based on photos." A post about Calgary home services said a website "needs to do four things in the first 5 seconds." Neither carried a source, because neither is sourceable. They are the same construction this article spends its opening section taking apart, published by us, on our own domain.

Both lines are corrected in the same change that published this post. We are naming them rather than quietly editing them, because an agency that debunks a statistic it is still running elsewhere on its own site has not really debunked anything.

Frequently asked questions

Is the eight-second attention-span statistic true?

There is no research behind the eight-second attention-span statistic. It is usually credited to a 2015 Microsoft Canada report, but that report carried it as a footnote to a statistics aggregator called Statistic Brain, which listed the National Center for Biotechnology Information and the Associated Press as its sources. When BBC World Service contacted both in 2017, neither could find any record of research supporting it. The goldfish comparison is backwards too: goldfish are used as a model system for studying memory precisely because they learn and remember well.

How much of a web page do people actually read?

No study has produced a general figure. Eyetracking can measure which words were fixated on particular pages in particular tasks, but those results do not add up to a single percentage. The widely quoted figures of 20% and 28% both come from Jakob Nielsen's own re-analysis, which calculated the most a reader could cover at an assumed 250 words per minute and then adjusted it down by judgement. Neither number appears in the peer-reviewed paper they are usually credited to.

Should I design my web page around the F-pattern?

No. Nielsen Norman Group, whose 2006 eyetracking produced the finding, argue in their own follow-up that the F-pattern appears when a page has large blocks of unformatted text and no strong cues, and that it hurts users and businesses because content on the right gets skipped. Their recommendation is to break the pattern with headings, subheadings, bold text, bullets and front-loaded opening sentences. A larger vendor analysis of 99 sites found no F-shaped bias at all, which is another reason not to build a layout around it.

Does everything important need to be above the fold?

No, and the sources cited for that rule say the opposite. Nielsen withdrew the original no-scrolling advice back in 1997, and Nielsen Norman Group state that people scroll when there is a reason to. Their 2018 eyetracking with 120 participants did find about 57% of viewing time above the fold on desktop, so the top of a page carries disproportionate weight, but attention declines with depth rather than stopping at the fold.

What is banner blindness, and does it only apply to ads?

Banner blindness is the tendency to miss elements sitting outside the structure someone is searching, and it is not limited to advertising. The researchers who named it in 1998 defined banners broadly and reproduced the effect with plain text that looked nothing like an ad, and a 2005 peer-reviewed replication found blank grey placeholder banners were fixated about as often as real ones. So your own content can be skipped when it sits in a boxed callout or a right-hand rail, which is why a primary action should never live only in a box.

Related reading

• Colour psychology, fact-checked, where the famous statistics also turned out to have no study behind them.
• Five marketing myths peer-reviewed research says are wrong, the first post in this series.
• Calgary fitness and gym marketing, one of the two posts corrected above.