We get a lot of inbound from teams who want their next site to score 90+ on Lighthouse. We always say yes, because the alternative is shipping a slow site and nobody is paying us to do that. But if Lighthouse 95 is the most ambitious thing on your performance roadmap, you are about a year behind.
Lighthouse measures one fast load on a clean cache from one location with no third-party scripts and no real users. It is a useful smoke test. It is not what your customers experience. It is not what Google ranks on. And it is not what your conversion team should be optimizing toward once you have shipped.
Performance is upstream of design, content, and conversion. We treat it the way Stripe treats latency: a product spec, not an engineering aspiration. Here is the 90-day plan we run on every site we ship.
Why Lighthouse alone is misleading
Lighthouse runs in a sterile environment. Real users hit your site on a six-year-old Android in a coffee shop. Real users have ad blockers, browser extensions, and slow DNS. Real users do not just load the homepage and leave. They navigate, fill out forms, and trigger interactive elements your synthetic test never touched.
A site can score 99 on Lighthouse and still have an eight-second time-to-interactive in the field, because the test does not measure the third-party tag manager the marketing team added last Tuesday. The synthetic number is green. The user is gone.
“Lighthouse is a smoke test. Real-user metrics are the report card.”
The three numbers that actually matter
We track three numbers per page, sourced from real user monitoring, not synthetic.
- 01LCP at the 75th percentile, segmented by device class and connection type. If your fast-Wi-Fi LCP is great but your 3G LCP is over four seconds, you are losing revenue from the fastest-growing segment of your audience.
- 02INP on every interactive element, specifically the buy button, the search bar, and the primary navigation. Slow interactivity correlates more cleanly with bounce than slow load does, and it is harder to spot.
- 03CLS during normal browsing flows, not just initial load. The biggest CLS offenders are usually third-party scripts that inject UI after the page is rendered. They will not show up in Lighthouse if they fire on user interaction.
What we strip out first
When we audit a site that scored well on Lighthouse but is bleeding conversion, the culprit is almost always the same. The marketing tag manager.
Heatmaps, session replay, A/B testing scripts, an Intercom chat widget, four different analytics packages, two retargeting pixels, a cookie consent provider that loads its own cookie consent provider, and a chat-bot that listens for your buy button. Each one is a few hundred kilobytes. Together they are a megabyte of JavaScript that runs on every page, blocks the main thread, and contributes nothing to the user experience.
We strip every one to start. Then we add back, one at a time, with measurement gates. If a tool cannot prove its conversion lift exceeds its performance cost in two weeks, it gets cut. We have killed nine-figure annual contracts on this rule.
“If a tool cannot prove its conversion lift exceeds its performance cost in two weeks, it gets cut.”
The 90-day plan
Day 0 to 30. Ship at Lighthouse 95+. This is the ground floor. We are not trying to win awards. We are clearing the table so we can see what is actually slow.
Day 30 to 60. Connect real-user monitoring on every page that converts. Web Vitals, custom interaction tracking, per-page LCP segmented by device. Build a dashboard the conversion team checks weekly. Make sure the marketing team has access too. They tend to be the source of the regressions.
Day 60 to 90. Prune. Cut every third-party script that cannot prove its lift. Test removing one platform-shipped feature per week (lazy-load images, defer animations, ship critical CSS only) and measure the impact on real users, not on Lighthouse.
What you actually own at day 90
At the end of ninety days you do not have a faster site. You have a site you can keep fast, because you know what is making it slow. You have a dashboard the conversion team owns. You have a removal policy for new third-party scripts. That is the difference between a Lighthouse score and a performance practice, and it is the only kind of speed that compounds.
Performance also feeds the work upstream. A site that loads fast for a crawler is a site an LLM will actually parse, which is the entry ticket to Citation Share. We wrote about why GEO does not replace SEO yet for anyone trying to connect the two.
Where this lands
Lighthouse 95 is the contract. Real-user metrics are the report card. The marketing team will keep wanting to add scripts. The right answer is not no. The right answer is prove the lift in two weeks or it does not stay.
If you want this kind of practice installed inside your own team rather than handed to an agency, the Build Your Own AI System page has the breakdown. The 90-day plan above is the same one we install. Same dashboards. Same removal policy. Same definition of done.
Scoping a build? See how we approach Websites & CRO.