TL;DR: Please take the speedPerception challenge to help us confirm results about which web performance metrics best match human perception of speed.
Last summer, I was involved in a study called “SpeedPerception”, a large-scale web performance crowdsourced study focused on the perceived loading performance of above-the-fold content aimed at understanding what “slow” and “fast” mean to users. I am now involved in the second part of this study which aims to confirm (or refute) our findings.
SpeedPerception: the general idea
Traditional web performance metrics, like those defined in W3C Navigating Timing draft specification focus on timing each process along the content delivery pipeline, such as Time to First Byte (TTFB) and Page Load Time. SpeedPerception’s goal is to tackle the web performance measurement challenge by looking at it from a different angle: one which puts user experience into focus by focusing on the visual perception of the page load process. We show the user sample video pairs of websites loading generated with http://www.webpagetest.org/ (WPT), and ask them which of the pair they perceive as having loaded faster.
In the first phase, we measured only Internet Retailer top-500 (IR500) sites in desktop size. Now we are testing whether the results we measured are true: in other words, do they only work for our IR500 sites on desktop? Will we get consistent results when testing Alexa top-1000 (Alex1000) homepages? Will we see the same results if we test on mobile size screens with mobile lie-fi performance?
In this second phase, we’re testing both mobile and desktop versions of both IR500 and Alexa1000 website home pages. We’ve also added a way of measuring the user’s time to click so we can compare apples to apples.
The goal is to create a free, open-source, benchmark dataset to advance the systematic study of how human end-users perceive the webpage loading process: the above-the-fold rendering in particular. Our belief (and hope) is that such a benchmark can provide a quantitative basis to compare different algorithms and spur computer scientists to make progress on helping quantify perceived webpage performance.
How was SpeedPerception created?
Videos were created using Patrick Meenan’s open-source WebPagetest (a.k.a WPT). We made 600+ videos of 2016 IR500 and Alexa-1000 home pages loading. The runs were done in February 2017. Videos were turned into gifs. Video pairs were grouped using a specific set of rules to help limit bias and randomness. Everything is available on GitHub at https://github.com/pdey/SpeedPerception.
Just like we did with the results of phase 1 of SpeedPerception, once the crowd-sourcing component generates a sufficient amount of user data, we will open source the dataset, making it available to the web performance community, along with the analysis of what we discover.
Please help us by taking the SpeedPerception Challenge now. Thanks.
Results from Phase 1
In phase 1, we discovered a combination of three values: an abbreviated SpeedIndex up to time to click (TTC) and an abbreviated Perceptual Speed Index up to TTC, in conjunction with startRender (or Render), can achieve upwards of 85%+ accuracy in explaining majority human A/B choices. Does the power of this new combination “model” hold true for all sites, or just our original data set? This is what we’re working on finding out.
If you’re interested in phase 1, here’s some more light reading:
- Perceived Performance of Webpages In the Wild Insights from Large-scale Crowdsourcing of Above-the-Fold QoE (PDF)