Evidence-based interface design

Interface design asks which choices in a web page or app screen change what people can do with it: how fast they finish a task, how many mistakes they make, how much they understand, and whether disabled users can use it at all. A small set of these choices has controlled experiments behind it: how long a line of text should be, how large a touch target must be, where an error message belongs, how much a page can show before people slow down. Many widely repeated rules rest on one lab study, a practitioner’s test of a few users, or a standard whose numbers were never checked against reading performance. Several rules that tested well in one setting reversed in another, and some famous rules, such as the three-click rule, failed when they were tested.

History

Two founding sets of numbers still shape practice. The contrast threshold in the Web Content Accessibility Guidelines (WCAG), the W3C’s standard for accessible web content, takes its 3:1 baseline from ISO 9241-3 and an ANSI standard for text and vision (W3C Web Accessibility Initiative, 2026b). Nielsen’s three response-time limits (0.1, 1 and 10 seconds) date from his 1993 book and cite Miller (1968) and Card et al. (1991) (Nielsen, 1993b). Both are examined under the controversies below.

A history of how the field moved from guideline-driven practice to controlled studies of specific design choices is not in the research.

Core concepts

What counts as a good outcome

The W3C Web Accessibility Initiative defines usability as being effective, efficient and satisfying, and accessibility as an equivalent user experience for people with disabilities (W3C Web Accessibility Initiative, 2016). ISO 9241-11:2018 likewise defines usability as effectiveness, efficiency and satisfaction in a specified context of use (RefHub (DriftSys), 2018). A systematic review of e-health usability questionnaires found that those instruments cover learnability, efficiency and satisfaction most often (Sousa & Lopez, 2017). Studies below therefore report completion time, error counts, success rates, eye fixations (a proxy for effort), comprehension, and sometimes only preference. These are different outcomes, and they often disagree: in one on-screen reading study, readers rated 55-character lines easiest, yet their reading speed did not follow their ratings (Dyson & Kipping, 1998).

Layout, structure and navigation

Visual complexity. In a 165-participant online study, higher visual complexity of websites reduced search efficiency and information recall (Baughan et al., 2020). The harm was not equal for everyone: people who preferred simple sites were hurt more by highly complex ones, and medium complexity was where all preference groups converged on speed. The complexity levels came from a computational model of perceived complexity, and the task rewarded speed, which the authors say limits generalisation (Baughan et al., 2020). Task difficulty matters too: in an eye-tracking study, on a simple task the most complex site was worst, but on a complex task a medium-complexity site was where performance peaked, so the penalty is not monotonic (Wang et al., 2014). A cross-cultural follow-up (65 Japanese and 84 US participants) found Japanese participants slower overall, with the gap growing as complexity rose; its hypothesis that culture-specific attention explains the gap was not supported (Baughan et al., 2021).

Link density and alignment. In web page mockups, alignment did not improve search performance (Parush et al., 2005). The abstract of that study, which ran two experiments (Hebrew and English), also reports poor performance on pages with many links and variable density and better performance with uniform density; it gives no effect sizes (Parush et al., 2005). In a study of white space, higher overall text density lengthened search (Software Usability Research Laboratory, 2004).

Menu position and clickability cues. In a supermarket-site experiment, a left vertical menu, a top fisheye menu and a top horizontal menu showed few significant differences; the authors attribute outcomes to affordances (what an element signals that it can do) rather than position (Murano & Oenga, 2012). Affordance is where a measured effect appears: in a randomised eye-tracking study, pages with weak clickability cues took about 22% longer and drew 25% more fixations than strong-cue versions of the same pages (Moran, 2017).

Where people look. Eye tracking of 120 participants shows viewing time concentrating at the top of the page: about 57% above the fold in 2018, down from about 80% in 2010. This describes attention, not task outcomes (Fessenden, 2018). Banners and ad-like content are often ignored (attention researchers call this banner blindness), covered under the controversies.

Typography and reading

Line length. A review of line-length studies reports that comprehension, reading speed and eye movements are affected by line length, and that font type and size, line and character spacing and screen structure also affect readability; it records a recommendation of no more than about 70 characters per line, since very short and very long lines both interrupt eye movements (Nanavati & Bias, 2005). On screen, 100-character lines were read faster than 25-character lines (Dyson & Kipping, 1998). In a controlled study on laptop, tablet and phone formats, 80 percent of maximum reading speed needed on average 13 characters per line for normally sighted readers and 8 for low-vision readers (Atilgan et al., 2020).

Size. A review finds the range of print sizes giving maximum reading speed spans a factor of 10, from about 0.2 to 2 degrees of x-height (Legge & Bigelow, 2011). A CHI 2016 study with 104 participants found readability and comprehension improved with font size up to 18 points, with weaker line-spacing effects (a secondary report) (Clark, 2023). The 18-point result recurs in other groups: for 28 readers with dyslexia, readability rose with size up to 18 points and line spacing had no significant effect, though the authors note the small sample and Spanish-only text (Rello et al., 2013); on an iPad, 90 UK and Thai participants read 18-point text faster than 14 or 16 point, with typeface not changing reading time, and age not changing the size effect (Chatrangsan & Petrie, 2019). Older adults prefer larger fonts, with a limit beyond which readability falls (Hou et al., 2022); in a Chinese-text experiment with older adults, larger size and spacing were preferred but tended to bring more errors and slower reading (Hou et al., 2018). A thesis found older UK and Thai adults preferred 16 point and younger adults 14 point or larger; this is preference, not performance (Kamollimsakul, 2016). Margins affected both reading speed and comprehension in one study, while line spacing did not affect reading performance but did influence preference (Chaparro et al., 2004).

Typeface. A systematic review of 42 empirical studies (2000-2025) finds serifs are not a significant legibility factor, familiarity of the typeface may matter, and no single size or typeface suits everyone (Ho Sang & Petrarca, 2025). Two studies found no benefit from specialised dyslexia fonts: across two experiments with 170 and 147 children, the Dyslexie font gave no faster or more accurate reading than Arial, and most children preferred Arial (Kuster et al., 2017); a single-subject study found no gain in rate or accuracy with OpenDyslexic against Arial and Times New Roman (Wery & Diliberto, 2016).

Colour and contrast

Reading is tolerant of reduced contrast for people with normal vision: in one study reading speed fell by less than half for a tenfold drop in contrast (Legge et al., 1987). Low-vision readers need more: in 16 of 19 low-vision observers the contrast at which reading rate halves was higher than in normal observers (Rubin & Legge, 1989). In a study of web colour combinations with 136 participants, higher contrast ratios generally gave greater readability but did not significantly affect retention (Hall & Hanna, 2004). Dark text on a light background gave better proofreading, with the advantage larger at small font sizes (Piepenbrock et al., 2014), and black on white was more legible than white on black in a glance-legibility experiment (Dobres et al., 2016). People with colour vision deficiency searched much worse for red targets in a real-world search task, so colour alone is a weak signal for them (Kugler et al., 2015). An experiment with 24 participants found colour-marked required fields led to fewer errors, faster completion and higher satisfaction than asterisks (Pauwels et al., 2009).

Forms

A controlled eye-tracking experiment (n=65) found that forms improved with 20 guidelines were completed faster, with fewer submission attempts and fewer eye movements, than the original company forms (Bargas-Avila & Seckler, 2014). In follow-up interviews, participants named format hints, error messages next to the field, and optional-versus-mandatory marking as most helpful, and labels above fields lowered fixation effort (Bargas-Avila & Seckler, 2014). An earlier controlled lab test of the same guideline set used 23 participants on real company forms (Seckler et al., 2013).

Error message placement and wording. An online study with 303 participants found error messages near the erroneous field performed better than messages at the top or bottom of the form (Seckler et al., 2012). Visual variants of an error message led to substantial improvements over the original in error comprehension and resolution in an online experiment (Khandwala & Dong, 2019). Baymard’s testing reports that specific error messages speed recovery and reduce the number of participants who give up (Scott, 2023).

Touch targets. A two-phase study of one-handed thumb use found speed improved with target size and recommended 9.2 mm for discrete and 9.6 mm for serial tasks (Parhi et al., 2006). A field dataset of about 120 million touch events measured error rates by target size, showed that touches are systematically skewed from the intended target, and reported that a compensation function cut the error rate by 7.79% (Henze et al., 2011).

Waiting and loading states

Progress indicators change how long a wait feels. In a 16-participant duration-matching test, a 5-second solid progress bar felt equivalent to a 5.61-second backwards-decelerating ribbed bar, which is 12.2% longer in real duration (p<0.001) (Harrison et al., 2010). The experiments in that study used 20, 10 and 16 young participants, so the result comes from small student samples (Harrison et al., 2010). Across four experiments with 68 students, more steps in a progress bar or throbber led to underestimated duration, with step effect sizes from small to large (partial eta squared 0.11 to 0.55) (Ziat et al., 2022). How placeholder screens compare is not settled; see below.

Accessibility for disabled users

Standards and what they cover. In usability tests with disabled users, WCAG 1.0 would have identified only 27% of the website accessibility problems found, and WCAG 2.0 added a marginal 5% (Romen & Svanaes, 2012). With 32 blind users on 16 sites, only about half of the problems users met were covered by WCAG 2.0 success criteria, and for the covered problems 16.7% of sites used recommended WCAG techniques that did not solve the problem (Power et al., 2012). W3C itself states that its cognitive-accessibility guidance covers barriers not in the normative WCAG 2.x text and is not needed for conformance (Seeman-Horwitz et al., 2021). Conformance is also unreliably measured: experienced evaluators did not reach 80% agreement when auditing WCAG 2.0 conformance (Brajnik et al., 2012). A user-based thesis found no significant relationship between most technical-guideline measures and user-based accessibility (Freire, 2012).

Does conformance help? A high level of conformance with WCAG 2.0 gave nondisabled users better task completion time and rate than low or very low conformance, while low and very low conditions did not differ on any outcome (Schmutz et al., 2016). A follow-up from the same group reports benefits for nondisabled and visually impaired users alike; it is not an independent replication (Schmutz et al., 2017).

Captions. Reviewing over 100 studies, Gernsbacher reports that captioning a video improves comprehension of, attention to and memory for the video, and notes that automatic captions still contain interfering errors (Gernsbacher, 2015). In a randomised experiment with second and third graders, captions helped children recognise and read more words and identify their meaning (Linebarger et al., 2010). A critic of the caption claim reports finding essentially no objective-measure research on same-language captions for neurotypical K-12 learners (Watson, 2025a). In a 209-learner experiment with English-as-a-foreign-language students, on-screen text duplicating narration helped instead of hurting, so redundancy effects depend on language familiarity (Toh et al., 2010).

Navigation aids for screen readers. Screen reader users completed two tasks in about half the time with a skip-link table of contents than with the screen reader alone, and automatically added ARIA landmarks (markup that labels page regions) shortened task times, though by less than the table of contents (Ferdous et al., 2021). A review page reports Watanabe (2007) found properly marked-up headings cut task time by 10% to 50%, with larger benefits for blind users (Watanabe, 2007). The same review page relays that Takagi et al. (2007) found adding heading tags and skip links did not significantly reduce reaching time or task-completion time for blind users (Takagi et al., 2007).

Alt text and labels. In an experiment with 18 sighted-simulated and 4 visually impaired participants, alt text on only important images gave better microtask performance than no alt text or alt text on all images (Zhong et al., 2020). Three blind screen reader testers completed 55.6% of 90 job-application attempts; unlabelled buttons and form fields and questions not tied to labels were among the comments describing critical blockers (Reuschel et al., 2023). This was an audit, not a comparison of labelled and unlabelled versions.

Automated checking. Automated tools find only part of the problems. Against a page with planted barriers, the best single tool found 37-41% of them and all tools combined about 71% (Duran, 2017). Nine tools run on 121 pages each found many issue instances that all the others missed (Pool, 2023). One evaluation found automated tests covered 23-50% of WCAG success criteria, with completeness from 14% to 38% (Vigo et al., 2013). A vendor reports its automated tests found 57.38% of total issues counted by issue volume, with colour contrast the largest category, and argues for counting by volume rather than by success criteria (Deque Systems, 2021).

Questions, controversies, and new developments

When should a form validate input?

The evidence points in opposite directions. In a 22-user test of six registration-form variants, the best inline-validation version beat a submit-time-validation control on completion time, errors, success and fixations, but the variant that validated before and while the user typed did worse than the other inline variants, so timing changes the result (Wroblewski, 2009). Two experiments (77 and 90 participants) found that showing the erroneous fields after the whole form was completed was best, and that immediate error feedback, which ISO recommends, performed worst (Bargas-Avila et al., 2007). A 2024 study of Arabic smartphone forms reports the opposite, that error messages before submission (immediate feedback) worked best (Alsalamen et al., 2024). Baymard’s usability testing reports that inline validation shortens error correction, but that validation firing before the user finished typing frustrated participants, and that it does not entirely replace the need for an error page (Scott, 2024). No single timing is supported across studies, and the strongest supported claim is that validating too early is harmful.

Does placement of labels and errors matter?

Evidence for label alignment is thin and conflicting. An informal eye-tracking test found right-aligned labels beside fields roughly halved fixations and completion time versus left-aligned, with labels above fields better in most cases (Penzo, 2006). A 312-participant eye-tracking study of three alignments found left-aligned labels were noticed fastest, yet top and infield alignments performed better than left in the fixation analysis, so it gives no single fastest alignment (Tangmanee & Ruchikachorn, 2018). Baymard’s study of 18 mobile e-commerce sites recommends labels above the field on mobile, with landscape orientation the exception; this is observation, not a timed comparison (Holst, 2013).

Error placement is also less settled than the 303-participant study suggests. An independent eye-tracking experiment with 32 participants and four error-message locations found no statistically significant difference in completion time, error recognition time, saccades or correction time between locations, though errors were spotted fastest when shown to the right of the field (Inal, 2016).

Which popular rules failed their tests?

The three-click rule was not supported: click count did not predict task success in 620 tasks by 44 users (Porter, 2003). The related advice to keep menus shallow is not simply supported either. A controlled experiment found a medium breadth/depth web structure beat the broadest shallow one, while increased depth hurt performance (Larson & Czerwinski, 1998); arranging links with more breadth and less depth gave shorter task times (Zaphiris, 2000). In a driving-simulator study of touchscreen menus, visual demand rose with depth for systematic reading but reached an optimum for search and memory tasks, so the best depth depends on the task (Louveton et al., 2024). Non-literate phone users were faster and more accurate with a 40-item multi-page list than with a four-level hierarchy (Medhi et al., 2013), and low textual literacy and low Raven’s scores predict worse navigation of hierarchies even in text-free interfaces (Medhi Thies et al., 2013).

Banner blindness also did not replicate cleanly. Benway’s result that extremely colourful, obvious banners tend to be ignored (Bayles, 2000) met Bachofer (1998), who found much higher banner recall and recognition; banner attention also depends on the task, since a group browsing aimlessly noticed banners better than a group searching for information (Pagendarm & Schaumburg, n.d.). Eye tracking with 26 participants in a practitioner study found that legitimate content with ad-like traits is also ignored (Pernice, 2018).

How long may a page make people wait?

Nielsen’s limits of 0.1, 1 and 10 seconds (Nielsen, 1993b) are criticised by a practitioner as never tested experimentally (Dubuc, 2018). Nah estimates tolerable waiting for information retrieval at about 2 seconds (Nah, 2004). A 196-student experiment with delays of 0 to 12 seconds found performance and behavioural intentions flatten from about 4 seconds of delay and attitudes from about 8 seconds; delay explained only about 2%, 5% and 7% of variance in performance, attitudes and intentions (Galletta et al., 2004). A 2021 mobile web-search study found delays mattered sizeably only beyond 7-10 seconds, with mobile users far more tolerant than the desktop figures reported earlier (Arapakis et al., 2021). Tolerance varies by group: participants from polychronic cultures were significantly less concerned with download delays than monochronic ones (Rose et al., 2003). In a video-search experiment, people terminated and avoided videos with longer download times, with a weaker effect when experienced download times were similar (Smith et al., 2019). Trade press reports a Google finding that 53% of mobile visitors leave a page that does not load within three seconds; the method is not described (Kirkpatrick, 2016).

Do skeleton screens beat spinners?

A Viget test with 136 participants, reported second-hand, found the skeleton screen produced the longest perceived wait (2.82 s) compared with a spinner (2.41 s) and a blank screen (2.29 s), and a slower later task (10.54 s against about 9.5 s) (The Plus Addons, n.d.). A thesis reports in two crowdsourced studies that a false-front page (a screenshot of the finished page that accepts clicks early) gave faster action completion and higher speed ratings than both a spinner and a skeleton-style placeholder, and no clear speed advantage for the skeleton over the spinner (Dufresne, 2026). The thesis also cites prior work that users prefer pages that load all at once to progressive loading (Dufresne, 2026).

Do attractive interfaces work better?

A 2019 meta-analysis (25 studies, 3,025 participants) found a small positive effect of interface aesthetics on objective performance (g = 0.12), with high heterogeneity and no significant moderator (Thielsch et al., 2019). A 2026 meta-analysis (31 studies, 18,794 participants) found a small to medium effect (g = 0.29), again with high unexplained heterogeneity (Schlamann et al., 2026); its record names interaction device and typography manipulation as significant moderators without level-by-level results (Schlamann et al., 2026). The effects on perception are larger. In an ATM-surrogate experiment, perceived aesthetics correlated strongly with perceived usability before and after use, and the system’s aesthetic level, not its actual usability, shaped post-use perceptions (Tractinsky et al., 2000).

Credibility follows the same pattern. Robins and Holmes found that the same content with a higher aesthetic treatment was judged more credible, within the first few seconds of viewing; the study used 20 library-science graduate students judging 42 screenshots, a convenience sample that may not generalise (Robins & Holmes, 2008). In 31 consumer health sites, visual design ratings were statistically significantly similar to credibility ratings (Robins et al., 2010). The Stanford study, in which 2,684 people rated two live sites each, found design look the most-mentioned feature in written comments (46.1 percent), which measures what people say rather than controlled design effects (Consumer Reports, 2002). A 2015 re-analysis of 15,750 comments from 2,041 participants found design affected 19 percent of evaluations of high-yield investment pages and 12 percent of other finance pages, far below the 46 percent, and reported that scam-like pages can deceive users mostly by their looks, an interpretation drawn from correlational comment data (Kakol & Nielek, 2015).

Are the WCAG contrast thresholds evidence-based?

WCAG’s 4.5:1 ratio is derived by multiplying 3:1 by a 1.5 contrast-sensitivity loss associated with 20/40 acuity (Arditi and Faye, 2004); it is a calculation, not a measured reading-performance threshold (W3C Web Accessibility Initiative, 2026b). The only reference given for that figure is a page-287 entry in a supplement of Optometry and Vision Science, not a full article (W3C Web Accessibility Initiative, 2026b). The developer of a rival method (APCA) states that WCAG 2 thresholds lack empirical testing and validation studies; this is an advocacy claim, with no studies cited on the page (Somers, 2022). A vendor blog supporting APCA likewise cites no human perception study: its evidence is theory, examples and a 500-pair test by the APCA author whose ground truth is not explained (Muth, 2022).

Does accessible design help everyone?

The “benefits everyone” claim rests on a thin set of studies. High WCAG conformance helped nondisabled users in one experiment (Schmutz et al., 2016), and captions have a large literature (Gernsbacher, 2015), yet task evidence for headings and skip links conflicts, and the 27% to 50% coverage figures show that conformance with the standard does not guarantee that disabled users can complete tasks (Romen & Svanaes, 2012; Power et al., 2012). See Inclusive design for the wider debate.

Broader connections

Interface design applies findings from perception, such as how contrast and size limit legibility, and from attention, such as banner blindness and fixation on the top of the page. The choices it measures are usually fixed into a design system of tokens and components; how teams reach them is the subject of Interface design process, and how problems are found in a finished design is covered in Usability evaluation. The same principle of fitting one display to a task governs visual representation of information, which covers charts and diagrams. Interfaces built by language models share these requirements and add defects of their own, described in AI-generated user interfaces.

Further reading

• Baughan et al. (Baughan et al., 2020): a controlled study of visual complexity and search, showing how user preference moderates the effect.
• Legge et al. (Legge et al., 1987): the classic measurement of how little reading speed depends on contrast for normal vision.
• Rømen and Svanæs (Romen & Svanaes, 2012): how much of what disabled users meet is covered by WCAG.
• Galletta et al. (Galletta et al., 2004): an experiment on delay thresholds, with small variance explained.
• Thielsch et al. (Thielsch et al., 2019): a meta-analysis of aesthetics and objective performance.
• Porter (Porter, 2003): a direct test of the three-click rule.