Performance as a design property: who the 75th percentile is

A performance target is a distributive decision, and the convention makes the choice visible. The numbers are judged at the 75th percentile rather than the median, so a target states how the slowest quarter of page loads must be treated (Google, 2024). Speed is a design property a reader feels before naming it. The strongest counter to this unit's thesis is that the convention is not arbitrary. Google (2024) holds that a small, stable, field-measurable set is what makes performance actionable, and that the metric lifecycle keeps the set honest. The units before it asked what a page must do for a reader. This one asks what it costs, and who pays.

The target is a claim about a population

The Core Web Vitals are three metrics: Largest Contentful Paint for loading, Interaction to Next Paint for responsiveness, and Cumulative Layout Shift for visual stability. Their thresholds are 2.5 seconds, 200 milliseconds and 0.1. A page passes only when all three are met at the 75th percentile, measured separately for mobile and desktop (Google, 2024). Two things follow. The unit of measurement is a population of page loads, not a page. A percentile also has a shape: the median describes the typical session, and the 75th describes how bad things may get for the slow quarter. The field survey makes that concrete: 365 kilobytes is the proposed budget for a load under three seconds at the 75th percentile (HTTP Archive, 2024).

Speed is perceived before it is measured

The percentile is objective. The experience it stands for is not. Bouch, Kuchinsky and Bhatti (2000) mapped objective quality of service to perceived quality in internet commerce, and they used focus groups to find the contextual factors that shape judgement. The psychological evidence points the same way, though it comes from outside the web. Witt and Sugovic (2010) showed in tennis and Pong experiments that performance and task ease change how fast a ball appears to travel. Perceived speed therefore depends on the perceiver's ability to act, not on the optical information alone. The perception was driven by context and expectation rather than latency alone, and it still correlated with objective measures. That is why a metric is a proxy rather than the experience itself.

Bhatti, Bouch and Kuchinsky (2000) estimated users' tolerance of service quality in commerce, and showed how those thresholds could be fed into server design. Their argument is that a user requirement can be translated into resource allocation, so the acceptable wait is set for an audience rather than a machine.

One metric set, one vendor, and the question of validation

None of that matters if the metrics are self-evidently correct. They are one framework with an owner. The Core Web Vitals are published by Google, and the field measurement it points to is the Chrome User Experience Report (Google, 2024). The earlier RAIL model, from the same publisher, sets different goals. It asks for a response within 100 milliseconds and an interactive load under five seconds on a mid-range mobile device over slow 3G (Google, n.d.). The two frameworks disagree, which is the point: thresholds are a convention that moves.

The research literature is more sceptical. Hossfeld, Metzger and Rossi (2018) examined the Speed Index, introduced by Google in 2012 to quantify how quickly the above-the-fold part of a page becomes visually complete. They argued that its relationship to web quality of experience remains far from clear. A de facto standard can still lack a demonstrated link to what people feel. Barakovic and Skorin-Kapov (2017) surveyed quality of experience modelling for web browsing, and they concluded that the quality-of-experience community has largely neglected the user-experience work. Together they pose the validation question: does the number track the person, or the machine?

The field data shows a heavier web with a longer tail

The Web Almanac measures what sites serve. In July 2025 the median desktop home page weighed 2,862 kilobytes and the median mobile home page 2,559 kilobytes, up from 2,652 and 2,311 kilobytes a year earlier. The median page made 77 requests on desktop and 72 on mobile, and JavaScript remained the most requested file type, at 1,291 kilobytes for the desktop page at the 75th percentile (HTTP Archive, 2025). Pages have grown again, and the growth still sits in the costliest resource to execute. The performance inequality gap holds: heavy pages fall hardest on users with cheaper devices and metered connections (HTTP Archive, 2024). Those users are not randomly distributed.

A desktop metric applied to a mobile tail

Kelton and colleagues (2020) argue existing page-load metrics were designed for desktops and ignore mobile interaction. They built muPLTest, a user-centred model for mobile: mobile users view several viewports during a load because they begin scrolling before it finishes. The slow quarter is disproportionately mobile, because the pass-or-fail metric was devised by asking a desktop question. Even on one device the experience has more than one shape. Salutari and colleagues (2020) crawled over 100 sites representative of the top million and crowdsourced 6,000 opinions on perceived page load time. They found the distribution multimodal with no more than three modes, and concluded that one metric cannot say whether a page loaded fast.

The mechanism, and what moves the number

Edgar (2024) explains that browsers do not wait for every file before rendering. Main text may appear while a logo or stylesheet arrives later, and elements already on screen then move to make room. That is the mechanism behind Cumulative Layout Shift, and it follows from a design decision rather than a slow network.

Work one shift. A hero image loads without width and height attributes, so the browser reserves no space for it. The headline below is painted first and then jumps down when the image arrives, which moves the line the reader was on. The page loaded quickly and the reader was still displaced. On the slow quarter, where the connection is metered and the device older, the same displacement lasts longer. That is why Cumulative Layout Shift is measured separately from load time.

Performance is one factor in whether people stay

The return literature is the qualification. Shiau and Luo (2013) surveyed 430 blog users, a hedonic system, and modelled continuance intention with expectation-confirmation theory extended by perceived enjoyment, habit and involvement. Continuance intention was predicted by involvement, satisfaction and perceived enjoyment, while habit showed no strong relationship. Speed is not the determinant, and Bouch, Kuchinsky and Bhatti (2000) found the same: judgement depends on expectation and context. A page that meets its 75th percentile may still be one people do not want to use. The evidence is about a blog rather than a page load, so it bounds the thesis rather than testing it.

Contested question: can one number carry the experience, and whose line is the 75th percentile?

Google (2024) publishes three numbers, one per facet, and requires all three at the 75th percentile, on the argument that a small, stable, field-measurable set makes performance actionable. Against that, Salutari and colleagues (2020) measured multimodal perceived load time and concluded that one metric cannot describe whether a page loaded fast. Barakovic and Skorin-Kapov (2017) argue the modelling effort needs several dimensions. Both sides accept that measurement is possible. They disagree about how much of the experience a number can carry. What would settle it is a study that records field metrics and user-reported experience in the same sessions on the same devices, and reports how much of the variation in reported experience the metric set explains, analysing the slow quartile separately.

The second dispute is distributive. The 75th percentile is itself a convention: stated in one vendor's guidance, measured largely in one vendor's browser, applied separately to mobile and desktop (Google, 2024). It is defensible, because an average hides the slow device while a percentile forces the slow quarter into view. The evidence on tolerance shows the acceptable wait to depend on context and expectation rather than a fixed constant (Bouch et al., 2000; Bhatti et al., 2000). What would settle it is a study that varies the threshold deliberately, holds the task constant, and measures task success and abandonment for the slowest quartile of devices, rather than the metric alone.

While the question is open, a designer should set the threshold as a policy and say whose loads it is protecting. Measure in the field, at the 75th percentile, and report the number separately for the cheapest devices the audience uses. Budget the heaviest resource first, because page weight is where the slow quarter is made.

Where the argument is heading

Three directions are visible in named work. The metric set is maintained as a set: Interaction to Next Paint was developed to address runtime performance more comprehensively than First Input Delay, and stable metrics change no more than once a year (Google, 2024). Field data is becoming public and comparative, which turns a site's performance into a statement about its users. The Web Almanac reports page weight by percentile and device, and it names the performance inequality gap (HTTP Archive, 2024). Optimisation is becoming a diagnostic loop rather than a budget, the direction Izripov (2026) sets out. The designer's task is not to hit a number once. It is to keep deciding, as audience and devices change, who the number is for.

The sources this unit rests on

• Google (2024). Web Vitals. web.dev. https://web.dev/articles/vitals The guidance sets the Core Web Vitals thresholds at 2.5 seconds, 200 milliseconds and 0.1 and says a page passes only when all three are met at the 75th percentile, separately for mobile and desktop.
• Google (n.d.). Measure performance with the RAIL model. web.dev. https://web.dev/articles/rail RAIL divides experience into response, animation, idle and load, setting a response within 100 milliseconds and an interactive load under five seconds on a mid-range mobile device over slow 3G. Dated 2020, it shows the thresholds are one convention among several.
• Bouch, A., Kuchinsky, A. and Bhatti, N. (2000). Quality is in the eye of the beholder. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/332040.332447 The authors mapped objective quality of service to perceived quality in internet commerce and used focus groups to find shaping factors. Perception depended on context and expectation rather than latency alone, yet remained correlatable with objective measures, the founding argument that performance is perceived rather than measured.
• Bhatti, N., Bouch, A. and Kuchinsky, A. (2000). Integrating user-perceived quality into Web server design. Computer Networks. doi:10.1016/s1389-1286(00)00087-6 The paper estimates users' tolerance of service quality in electronic commerce and shows how those thresholds can inform server design. It carries the bridge from user judgement to resource allocation that the argument depends on.
• Hossfeld, T., Metzger, F. and Rossi, D. (2018). Speed Index: Relating the Industrial Standard for User Perceived Web Performance to web QoE. QoMEX 2018. doi:10.1109/qomex.2018.8463430 The authors examine the Speed Index, introduced in 2012 to quantify the visual completeness of the above-the-fold page, and argue that its relationship to web quality of experience remains far from clear. It supplies the validation question for any performance metric.
• Barakovic, S. and Skorin-Kapov, L. (2017). Survey of research on Quality of Experience modelling for web browsing. Quality and User Experience. doi:10.1007/s41233-017-0009-2 The survey covers system, context and human influence factors on browsing and concludes that the quality-of-experience community has neglected the relevant user-experience work, and argues for multidimensional models that qualify a single metric set.
• Kelton, C., Ryoo, J., Balasubramanian, A., Bi, X. and Das, S. R. (2020). Modelling User-Centred Page Load Time for Smartphones. MobileHCI 2020. doi:10.1145/3379503.3403565 The authors argue existing page-load metrics were designed for desktops and ignore mobile interaction, and present muPLTest, a user-centred model for mobile, observing that users view several viewports during a load because they begin scrolling before it finishes.
• Salutari, F., Hora, D. d., Varvello, M., Teixeira, R., Christophides, V. and Rossi, D. (2020). Implications of the Multi-Modality of User Perceived Page Load Time. MedComNet 2020. doi:10.1109/medcomnet49392.2020.9191615 The study crawled over 100 top-million sites and crowdsourced 6,000 opinions on perceived load time. It finds the distribution multimodal with no more than three modes and argues a single metric cannot assess load speed.
• Izripov, Y. Z. (2026). Determinants of Core Web Vitals metrics in high-load web platforms. Soft Measurements and Computing. doi:10.36871/26189976.2026.07.001 The article develops a multi-factor model for First Contentful Paint, Largest Contentful Paint, Time to First Byte and Time to Interactive, separating server, network, resource and client factors. It proposes a cycle of measurement, diagnosis, action and validation, treating performance as a process.
• Edgar, M. (2024). Cumulative Layout Shift. In Speed Metrics Guide. doi:10.1007/979-8-8688-0155-6_9 The chapter explains that browsers begin rendering before all files arrive, so text can appear first while a logo or stylesheet arrives later and rendered elements then move. It supplies the mechanism behind a page that shifts.
• Witt, J. K. and Sugovic, M. (2010). Performance and Ease Influence Perceived Speed. Perception, 39(8). doi:10.1068/p6699 Tennis players judged a ball faster on trials they hit out of bounds, and participants judged virtual balls slower when playing Pong with a bigger paddle, so perceived speed depends on the ability to act rather than on optical information alone. The finding is about perception rather than about the web, and it grounds the unit's claim that a metric is a proxy.
• Shiau, W. and Luo, M. M. (2013). Continuance intention of blog users: the impact of perceived enjoyment, habit, user involvement and blogging time. Behaviour and Information Technology. doi:10.1080/0144929x.2012.671851 The authors surveyed 430 blog users and modelled continuance intention by structural equation modelling. It was predicted by involvement, satisfaction and perceived enjoyment while habit showed no strong relationship, showing that return is driven by more than speed.
• HTTP Archive (2024). The Web Almanac 2024, Page Weight chapter. https://almanac.httparchive.org/en/2024/page-weight The chapter reports the proposed budget for a load under three seconds at the 75th percentile, 365 kilobytes, and names the performance inequality gap, in which heavy pages fall hardest on users with cheaper devices. It supplies the proposed figure and the name of the gap.
• HTTP Archive (2025). The Web Almanac 2025, Page Weight chapter. https://almanac.httparchive.org/en/2025/page-weight The chapter reports that the median desktop home page weighed 2,862 kilobytes in July 2025 and the median mobile home page 2,559, up 7.3 and 8.4 per cent, and that JavaScript remained the most requested file type.