The 90th percentile doesn't lie
An average load time is a reassuring number. It is also a liar. A nice average hides exactly the people you should be worrying about: the visitor on a slower device, with a so-so connection, on the train between Groningen and Utrecht. They barely move the average, but they are still someone who wants to buy something.
That is why we do not look at the average, we look at the 90th percentile (p90). In plain terms: how fast is your webshop for the slowest 10% of your visitors? Score well there and you know for certain the other 90% have it at least as good. It is the strictest yardstick there is, which is exactly why it is the only one that really means anything.
For Haibu, that p90 now sits consistently around 90 out of 100. In other words: even your slowest visitors get a webshop that feels fast. Here is how we got there, and the role RUMvision 2.0 played.
Why p90, and why RUMvision 2.0 keeps hammering on it
RUMvision launched version 2.0 in April 2026, a complete rebuild of their Real User Monitoring platform. RUM means you measure the actual visitor, not a simulated test. Or, as they put it themselves: lab data is a simulation, RUM is reality.
The best thing about 2.0 is that it explicitly pushes you toward that 90th percentile. A PageSpeed score or a green Core Web Vitals check is often fine for the average visitor, while a whole group of people are having a slow experience without you knowing about it. RUMvision 2.0 puts that group front and centre, including a Health Check that automatically detects issues like layout shifts (CLS), slow interactions and failed bfcache hits, and closes them again as soon as they are resolved.
So you can no longer hide behind an average. As it should be.
The case: Haibu from 70 to 90
Haibu runs on Magento 2 with a Hyva frontend, a fast foundation to begin with. But already fast is not a finish line for us, it is a starting point.
The RUMvision graph over the past year tells the whole story at a glance. Until the end of September 2025, the p90 UX score (and the FUX score, which specifically measures the first page load) hovered around 70. Fine, but no more than that. Around 30 September 2025 the line makes a clear jump into the low 80s, and then climbs steadily to around 90, where it has stayed ever since.
No coincidence, no measurement glitch: that is the moment a series of performance optimisations went live.
What exactly we did
Re-engineered Varnish (the biggest lever). We rewrote the entire Varnish configuration using our own elgentos/magento2-varnish-extended, including Xkey for smarter cache invalidation and targeted caching of media and static files. The result: a higher cache hit rate, a lower Time To First Byte (TTFB) and therefore a faster Largest Contentful Paint (LCP). This is by far the main reason for the jump at the end of September. More on that approach in our Magento 2 Varnish taskforce.
Cleaned up the checkout success page. We removed a redundant database query from the success page (plus 60-odd lines of dead code), so the post-purchase page renders faster.
Trimmed the Prismic integration. The content integration was loading more products than necessary; we now filter only the relevant ones, making queries lighter.
A snappier cart. We sped up the cart interactions so changes give immediate feedback. Better for Interaction to Next Paint (INP) and for how it feels.
None of it is rocket science. Together they lift the p90 by some 20 points, and that is exactly the point: measurable results for real visitors, not for a test server.
Why this would have been guesswork without RUM
This is where the whole point of RUM comes together. Had we measured these optimisations with PageSpeed Insights, we would have had a nice lab number and a pat on the back. But we would not have known whether the visitor on that slower device noticed anything.
With RUMvision 2.0 we do see it: the p90 creeping upward, day after day, based on thousands of real sessions. And since v2 we tie our deploys to it via annotations, so every release shows up as a vertical line on the timeline. If a score drops or climbs, we immediately see which release it belongs to. No after-the-fact reconstruction, just the facts, lined up.
Two Groningen companies, one obsession with speed
There is a local edge to our enthusiasm here too: RUMvision, like us, is from Groningen. Two northern companies working on the performance of Dutch and international webshops. No nonsense, just getting it done, and letting the numbers do the talking.
Want to know where your webshop stands at the 90th percentile, for the visitors you probably cannot see right now? We are happy to take a look. Get in touch and we will put your shop's real numbers on the table.
