What your website audit tool can't see

A slow page that was already "optimized"

The first example that comes to my mind is a community website for architects and interior designers. It's a platform similar to Pinterest, with interior design projects. Another team had already optimized it: They installed a cache and made a dozen changes, in order to make it quicker. Still, when we started to work on it in September 2026, the page with the projects took 6.1 seconds to answer from the server. The reason: The projects were shown in a new random order for every visitor, so the page was excluded from the cache and built from scratch on every visit.

We changed the order so it rotates once a day, not on every visit. Every project still gets its turn at the top, but now the page can be cached. We measured the same metric twice on Sept. 23, 2026: the server response time, meaning the time until the first byte of the page reaches the browser. Before the change, uncached: 6.1 seconds. After the change, served from the cache: 0.06 seconds. The cache did the speed-up; the change in the order is what let the page into the cache.

They had optimized it, but they hadn't looked at the real usage—the daily usage on phones, not only the lab test. What we did wasn't only optimizing, it was a change of perspective. No score tells you that a random order keeps your page out of the cache.

Our own website

On our own website, when we made our audit on Aug. 1, 2026, we had a really low mobile score of 67 in Google Lighthouse and a Largest Contentful Paint (LCP), the time until the main content shows up, of eight seconds, the average of four tests. Our results varied between 3.3 and 9.3 seconds. With the advertising pixels blocked, we had a score of 81 and an LCP of 3.7 seconds. The pixels weighed around 427 kilobytes.

In the lab, the scores were showing bad results, but in real daily usage the site was okay. In our experience, there is often a difference between the lab score and the real field, because Lighthouse simulates a phone on a slow connection, and many visitors have a faster one.

We ran more tests, and we found that the cookie plugin was active but wasn't blocking the pixels. Once we discovered that, we made the change, and after the fix the pixels waited for consent in 10 out of 10 test runs.

Why we tested 100 websites

In the speed audits we run for our clients, we felt there was a big difference between the lab score and the real-life experience. That's why we ran our own research, The Website Gap, published by Clan of Creators in October 2026, to confirm our hunch, or to learn from it.

We tested 100 business-to-business (B2B) software-as-a-service (SaaS) websites, 50 in the U.S. and 50 in the U.K. We took companies with roughly 50 to 1,000 employees from public growth lists, like the Deloitte Technology Fast 500 and the Sunday Times Tech Track, wrote our criteria before we started, and picked at random among the companies that met them. None of them are our clients. We collected all the data on Oct. 10 and 11, 2026: the homepage tests, the network checks and the questions to the artificial intelligence (AI) assistants. We tested each homepage once in Lighthouse, on a simulated phone, and took the real-user data from Google's Chrome User Experience Report, which covers the 28 days before collection. The full method is in the report.

Speed: the lab and the visitor

One thing to know before the numbers: the lab test measures only the homepage, while the Chrome data from real users covers the whole domain, all the pages together. So the two are not directly comparable. They are two different views of the same website.

In our research, the median homepage took 12.2 seconds to show its main content in the lab. For real Chrome users on phones, the median was 2.2 seconds across the whole domain. In the lab, 88 of 99 homepages had a "poor" Largest Contentful Paint result, over four seconds. In the real-user data, 63 of 97 websites had a "good" Largest Contentful Paint result, under 2.5 seconds for three out of four visits. That is Google's rating for this one metric, not a verdict on the websites overall.

So for many of these websites, the lab score makes them look slower than they feel to real visitors. Loading time is only one part of the experience, though. A fast page can still confuse a visitor or lose them somewhere else.

At the end of the day, we need to find the balance between the lab score, which scans the scripts, the pictures and everything, and the real usage: how the website works for a normal user, not only in the lab.

Data before the banner

In our research, 39 of 100 homepages sent data to LinkedIn or Meta before anyone touched the cookie banner, and 31 of them had a consent platform installed.

We saw the same problem closer to home. On our own website, the banner was there, but the pixels didn't wait for it. On the community website above, there was no banner at all. On 31 of the 39 SaaS websites, the banner was there, but the pixel didn't wait for any approval or refusal. The pixel started right after you landed on the website.

The two-minute test: Open a private window, like an incognito window. Open the inspect tool in Chrome and go to the network tab. Don't touch the banner, and you can easily see what data is leaving the website you want to test.

We report what the browser sent, not whether it is lawful.

The first line

In our research, only 33 of 98 homepage headlines (the H1, the main heading of the page) name the product type, like "contract management software." The other 65 don't. The page title, the one in the browser tab, names the product type on 68 of 98 websites. The information exists, but it's missing from the H1. And 55 of 98 say in the first lines who the product is for.

In my experience, teams often choose a slogan for the headline to sound premium, instead of being practical and straight to the point. But we are living in a world where being practical and straight to the point saves time. We need a mind shift to understand that being practical is not bad. It can be good, because you are saving time for your users.

Artificial intelligence and the way you name what you do

In our research, when we asked AI assistants about each company's exact niche, at least one of them recommended 93 of the 100 companies. When we asked about the broad category, only 31. On the niche questions, ChatGPT mentioned 83 companies, Gemini 84 and Perplexity 63.

This covers our questions and these 100 companies, not every product or every way of asking. But the pattern was clear: The assistants found most of these companies when the question was specific, and missed most of them when it was broad. Companies need to think about their target, the people they want to reach, and the words those people use when they search.

Asked the same question twice, ChatGPT changed its answer for 30 of 100 companies. So is ChatGPT recommending you? One check is a sample, not a measurement. Ask twice before you draw a conclusion.

What the tools catch, and what they miss

Coming back to the community website: On Sept. 23, 2026, before we started, its homepage had a Lighthouse mobile score of 69, and the main content showed up after 13.6 seconds on a simulated phone. After our changes, measured the same way on the same day, the score went up to 98 and the main content showed up after 2.2 seconds.

But there is a catch. There is a part the tools don't really catch, or that is hard to find. In the season when they opened registrations for their yearly projects contest, a new version of PHP, the programming language the site runs on, broke the forms: contest registrations, project uploads and login returned errors. The audit tool didn't catch it. The audit tool isn't looking at the big picture. It only looks at the lab test and at what files are sent or loaded. It's not looking at the entire server.

That's why it's great to use Lighthouse to find the heavy scripts and so on, but these tools measure what's easy to measure. A person sees the real cost, how each thing is used, how to fix it and what needs to be fixed first, because it's also a perception thing.

One hour this month

If you have one hour this month, check these by hand on your own website:

• Test your lead forms once a month. Sometimes a version changes, a plugin comes in, an update, and so on.

• Open your website in incognito mode, open the inspect tool, go to the network tab and see what your site is really sending, and where.

• Check the speed in Search Console before the Lighthouse score (review the Core Web Vitals report).

• Read your H1 without context. Would a stranger know what you are selling, based only on the H1?

• Search for your niche with the words your buyers use to find software or tools like yours, and try it twice to see how the assistants really react.

• Look at the review websites in your category. In our research, 74% of Perplexity's recommendation answers linked to review and comparison websites like G2 and Capterra. That was an early, exploratory read, and it doesn't prove that a profile there gets you recommended. It only shows where the assistants were looking when they answered.

One thing to remember

Check not only with tools: Check the real experience, the human experience on the website. It's not all about tools. It's also about how it feels and how it lands on the end user's screen.