SEO · 05
Web Performance Consulting.
Your site gets slower release by release, and one lab score does not show it. I measure what your users get, set budgets per page type, and check them on every release.
You hire me when
- Core Web Vitals are red in Search Console.
- Your interactive pages fail INP.
- Trackers and tag managers ate your speed.
- PageSpeed scores jump and nobody trusts them.
- Speed is a project every two years, not a release gate.
What you get
- Field data first: Search Console and Chrome's user data per page type, then lab runs, repeated.
- Budgets per page type: HTML weight, server response, INP.
- The audit as tickets, every to-do with a pass rule.
- Third-party scripts moved behind the first user interaction.
- Compression and CDN settings checked, with a check on every deploy.
- QA training, so every release is measured before and after.
- Our INP whitepaper for your developers.
Why speed slips back after every project
Speed is not a state your site reaches once. Every release adds something: a tracker, a framework update, a bigger image, a new widget. Nobody adds a second on purpose.
Google measures what your real visitors get, not what your developer’s laptop gets. On page experience: “Core Web Vitals are used by our ranking systems.” (Google Search Central, Understanding page experience)
So a speed project that ends with a green score and no budget is gone again two releases later. My goal: speed is a release gate, not a project every two years. Every page type has a budget, and every release is measured against it.
What slows sites down in 2026
Big images and render-blocking CSS still happen. They are rarely what keeps a site slow. These are the failures I measured in my audits, and the fix I specified for each:
| What breaks | What I measured | What I specify |
|---|---|---|
| Pages too heavy for what they show | 1 percent text in a 900 KB list page; trimmed to about 200 KB, no visible change | 150 KB of unzipped HTML, 250 KB at most |
| Compression that makes pages bigger | Brotli bigger than gzip | One CDN setting, one check per deploy |
| Scripts eat what the fixes won | Every task marked done, the speed score went down | A pass rule per task, third parties after the first interaction |
| Deploys delete files the page still needs | 95 percent of old JS and CSS files gone; each answered with a 500 KB start page | Hashed file names, old files kept for months, a short plain 404 |
| One lab run taken as the truth | 10 runs of one page: under 40 to over 90 | Field data first, lab runs repeated, the median per page type |
Pages too heavy for what they show
A list page shipped about 900 KB of HTML. 1 percent of it was text. The rest was framework data, inline CSS and inline icons. Almost every page type of that site was over 250 KB.
- I trimmed a real page without changing what it shows: from about 900 KB to about 200 KB.
- On another site, a lean rebuild of the same list page went from about 1.5 MB to about 120 KB, and its Lighthouse score from under 30 to over 90.
The phone has to download, parse and hold all of it, before and after it is compressed. As I put it in a 2026 workshop: compression hides the weight from the network, not from the device. The details of both cases are on my technical SEO page.
What I specify: 150 KB of unzipped HTML per page, 250 KB at most, checked per page type on every release.
Compression that makes pages bigger
Brotli should be smaller than gzip. On one site it was bigger, page type after page type.
Recompressing the same page at any Brotli level beat the live file. The cause was a streaming setting at the CDN that compressed the page in tiny pieces. Set right, the page is much smaller than what was served. The full case.
What I specify: one CDN setting, no template change, and a check on every deploy: Brotli bytes smaller than gzip bytes. I run the same check on this site, see below.
Scripts eat what the fixes won
One team implemented almost every task from its workshop with me. I re-measured its core page types on three dates over three months.
- The speed score never reached the goal. It went down, not up.
- The main content loaded faster, but the page spent longer blocked by scripts, so the score fell anyway.
- Nobody owned the question whether the result arrived on the live page and stayed there.
This is not new. In an older audit, analytics started only after over 3 seconds, because one script chunk loaded the next. The page measured had a Speed Index of almost 5 seconds, against a goal of 1 second.
What I specify:
- Every task gets a pass rule on the live page, not a checkbox.
- Conversion, ad and other third-party tags load after the first user interaction, at the earliest. How to do that with Google Tag Manager is in my article on loading third parties later (also in German).
- The next click can be faster too: prefetch the pages behind the visible links.
Deploys delete files the page still needs
A page that is cached in a browser, a CDN or Google’s index points to the CSS and JavaScript of the release it was built with. Delete those files on deploy, and the page breaks or loads slowly until every copy is refreshed.
- Of about 2.5k old JS and CSS files, 95 percent answered 404. One file went from 200 to 404 within an hour of a deploy.
- Every missing file answered with a 500 KB copy of the start page instead of a short 404.
- On another site, fingerprinted JavaScript URLs answered with a 500 error after each build.
What I specify: a content hash in every file name, old build files kept online for months, and a short, plain 404 for a file that is really gone.
One lab run taken as the truth
- 10 PageSpeed runs of the same page scored from under 40 to over 90.
- In the same audit, three runs on one day scored about 50 to 75.
Teams fix, run PageSpeed once, and celebrate or panic. Neither means anything on one run.
Google’s Core Web Vitals report uses a different source: “The CrUX report gathers anonymized metrics about performance times from actual users visiting your URL” (Search Console Help, Core Web Vitals report).
What I specify: field data first, per page type. Lab runs repeated, and the median compared, before and after each release. A single score is never a ticket.
How I work
- Field data first. Search Console’s Core Web Vitals report and Chrome’s user data per page type, before any lab tool. That shows which page types are slow for real people, on real phones.
- The same page, measured many times. One PageSpeed run is not a number. In one 2026 audit, 10 runs of one page scored from under 40 to over 90, so I compare medians, never single runs.
- INP traced to the interaction. PageSpeed and Search Console show that you have an INP problem. Which tap causes it takes analysis, with the attribution build of the web-vitals library. Holger Guggi and I wrote this down in a 44-page INP whitepaper; in 2024 I wrote an INP guide for one client’s developers.
- Budgets per page type. HTML weight, server response and the field metrics. The budgets I set in my audits: 200 ms server response and 150 KB of unzipped HTML, plus Google’s own threshold for a good INP, 200 ms.
- The audit as tickets. Every finding ends as a to-do with a pass rule. I re-check the list until every item passes.
- QA owns it after me. I train the QA team on the performance tools, so every release is measured before and after it goes live, without me.
This site, measured the same way
This site is static HTML on Cloudflare Pages since 21 October 2026. On 12 November 2026 I fetched every URL in its sitemap again, once with each compression.
- 109 URLs, every one answers 200.
- Median about 43 KB of unzipped HTML. One page breaks my own 150 KB budget: the workshop list, at about 190 KB. None is over 250 KB.
- Brotli is bigger than gzip on about a quarter of the pages, and only about 1 percent smaller over all of them. On 24 October it was a handful of pages. My own CDN shows the failure I find in audits, and in three weeks it grew. The cause is the on-the-fly compression at the CDN. The fix is in my build: every page precompressed at Brotli’s highest level, 18 percent smaller than gzip in total, served by a small function. On the deployment host it already works; on www it waits for one CDN setting.
- Build files carry a content hash in their names. A missing one answers a plain 404 of about 11 KB, not the start page.
- One third-party script: Google Tag Manager, loaded async.
- PageSpeed Insights, mobile, on three pages on 24 October 2026: performance 95 to 98, lab LCP 2.4 to 2.6 seconds. On 12 November the PageSpeed API refused my requests with a quota error, so these numbers are from October.
Run PageSpeed Insights on this page and check my numbers.
Questions I get in every performance audit
Does speed affect our Google rankings?
Yes, but it is not the main thing. Google: “Google Search always seeks to show the most relevant content, even if the page experience is sub-par.” (source) And: “We highly recommend site owners achieve good Core Web Vitals for success with Search and to ensure a great user experience generally.” (source) Speed wins when your content is as good as the next result.
Do we need a PageSpeed score of 100?
No. Google: “trying to get a perfect score just for SEO reasons may not be the best use of your time” (source). My rule for the lab score is at least 80 on mobile, better 90, as a budget, not a trophy.
What is a good INP?
Google: “Good INP values are 200 milliseconds or less. Poor values are greater than 500 milliseconds.” (web.dev, Interaction to Next Paint) In our whitepaper’s terms: every interaction has three phases, and JavaScript causes two of them, the input delay and the processing time.
We fixed it. Why is Search Console still red?
The report shows field data from real visits, so it trails your release. When you validate a fix, Search Console says: “Click Start Tracking to start a 28-day monitoring session” (source). Plan for about four weeks before you judge it.
Do we have to remove our trackers?
Rarely. Most of them can load after the first user interaction instead of during the page load. How to set that up in Google Tag Manager.
What do our developers get?
Budgets per page type, tickets with a pass rule, the INP whitepaper, and a check they run on every release.
Sources for the Google quotes on this page, read on 2026-11-12: Understanding page experience and Understanding Core Web Vitals, Google Search Central; Core Web Vitals report, Search Console Help; Interaction to Next Paint, web.dev. The client numbers come from my own audits, workshops and handovers.
From real jobs
- 2026: one list page trimmed from about 900 KB to about 200 KB of HTML, with no visible change.
- 2026: Brotli bigger than gzip; one CDN setting was the cause.
- 2026: ten PageSpeed runs of one page, from under 40 to over 90.
- 2024: an INP audit and an INP guide for the developers.
- 2023: INP workshops for a platform's developer teams; our 44-page INP whitepaper.
- A QA team trained on performance tools, so speed is checked on every release.
- About 35 performance audits with to-do lists.
Teams like yours
Some of our customersNews publishers
Dailies, agencies, broadcasters: Google News, Discover, training.
ExamplesOnline shops
Webshops and marketplaces: category pages, product pages, speed.
ExamplesMarketplaces and job platforms
Real estate, jobs, cars: list pages decide traffic.
ExamplesFintech, finance and insurance
Banks, insurers, lenders, fintechs: SEO under regulation.
ExamplesSomething else in the way?
How consulting works- External SEO lead
- Relaunch and migration SEO
- Technical SEO
- International SEO
- SEO for newsrooms
- Content strategy and search demand
- Search data and traffic forensics
- SEO agency cleanup
- GEO and AI search
- Vibe coding training
- AI automation consulting
- AI-first transformation
- Custom tools for one job
- Design thinking
- Growth
- Growth strategy analysis
- SEO workshops and training
- Keynotes and talks
- Sparring and second opinion
- Dev, product and marketing mediator