Measured cold every week. Fixed with a way back.
Each week an agent opens your site the way a first-time visitor meets it, with nothing cached, and writes down what slows it. The changes that can be undone it makes itself, each with a rollback point; the ones that need judgment go to an engineer. Every figure comes from a measurement, and no score is promised in advance.
Audited on Monday, fixed on Tuesday, proven on Friday.
A sample week for a shop site with four layouts. The agent does the measuring and the reversible work; the engineer takes what calls for a decision.
The cold-cache audit
Each layout is loaded three times on mobile and three times on desktop, the cache emptied before every run, so each figure is a first visit. A warm run never counts.
The fix plan, with the saving per item
Every item carries the saving it should bring, taken from the audit figures. Weight added since the last run, a new script or a plugin that grew, is flagged at the top.
Safe fixes, each with a rollback point
Images, caching, compression and headers are put right, and last week’s fixes go back in if an update undid them. Originals are renamed, never deleted.
Re-measured, and the engineer’s list
The same cold runs are repeated, so before and after compare like with like. An engineer reads the report and takes the templates, critical styles and third-party calls.
Four jobs a script can do safely, done every time.
The agent measures, weighs, plans and applies only what can be reversed. Each run leaves a written log with the figures behind it.
Cold-cache audit
Three runs per layout on mobile and desktop, with nothing held in the cache. The figures describe a visitor arriving for the first time.
Resource breakdown
Every script, stylesheet, font and image is weighed, and anything loaded twice or never used is named. One audit of a live platform found its homepage script downloaded twice, three icon sprite files on a single page and a first visit of about 2 MB, while the server itself answered quickly. The full report is shared on request.
Fix plan with measured impact
A ranked list, each item with the saving it should bring, worked out from the measurements instead of a guess.
Safe fixes with a rollback note
Images converted to WebP or AVIF, minification, caching, compression, headers and edge rules, each written up with a rollback note per file.
Four things the agent leaves to a person.
A change made automatically in any of these is how sites come back broken, so the line between agent and engineer is drawn in writing.
Template rewrites
Wasteful markup, an order of loading that holds up rendering, and structure at the level of the layout.
Critical styles per layout
Above-the-fold styles placed inline, a strategy for loading fonts, and fonts cut down to the characters in use.
Third-party decisions
Whether each tag, tracker and chat widget is worth what it costs to load. That is a business decision, taken with you.
Review before close
A person reads every report and every monthly run before it is sent on to you.
Two lanes, closed by one measurement.
The agent audits, plans, applies the safe fixes and keeps a rollback point. The engineer takes the templates, the critical styles and the third-party calls. Both lanes finish at the same cold re-measure, and the weekly re-audit catches anything that creeps back.
Figures you can check for yourself.
A speed report taken on a warm cache flatters the site. Each figure here is tied to the run that produced it, recorded the way your next visitor arrives.
Cold cache, three runs, never warm
A warm cache hides most of what a new visitor waits through, so no figure in the report comes from one.
Before and after on the same basis
The proof is the identical cold run, repeated once the fixes are in place.
Rename, never delete
Changed files keep their originals under a new name, so any change comes undone in minutes.
A real browser on your real pages
Findings come from loading your own pages in a browser, never from a third-party database of scores.
The audit runs anywhere; the fixes go as far as the platform allows.
Each platform sets how deep we can reach, so the proposal says in advance which fixes apply to yours.
Static HTML and custom PHP
Everything, through an FTP account limited to the web folder. No hosting login is asked for.
WordPress
Everything inside the site. Server-level changes only where the audit shows they would count.
Shopify
Theme level only: app scripts, render order, lazy loading. Images and caching belong to the platform; when the catalogue is the slow part, the Shopify Catalogue Agent takes on the products.
React / Next.js
Build-level fixes, delivered as pull requests to your repository.
Wix, Squarespace, Webflow
An audit and advice on settings. We sell no fix the platform will not let us apply.
Any public site
The audit itself runs on any address the public can open.
Four things you always get.
Scope is counted in unique page layouts, not pages: pages sharing one design and set of assets count once, the agent’s crawl does the counting, and the proposal states the number. A hand-coded site where every page differs is scoped on its own terms. The edge and font layer can be added to any engagement. Weekly monitoring re-audits cold, alerts on a regression such as a new script or a plugin that grew, puts the safe fixes back and sends a monthly report an engineer has read; it runs month to month on one month’s notice. Agencies and freelancers can take the work unbranded: we never contact your client, and resale terms are agreed in writing.
A written brief
What we agreed to build, in one document.
A fixed price
Set in the proposal, before work starts.
Staging + approval
You see it and approve it before it goes live.
Documentation
How it works and how to change it, handed over.
Before the agent measures your site.
Send any other question in writing; a written answer follows within one business day.
What does the agent do alone, and what does it not?
Unattended, it measures cold, weighs every resource, writes the fix plan and applies the safe fixes: image conversion, minification, caching, compression, headers and edge rules, each with a rollback note. It never works alone on templates, critical styles or third-party scripts. Those belong to the engineer, and a person reads every report before you see it.
What if a safe fix breaks something?
It is rolled back. Each change has its rollback note and originals are renamed rather than deleted, so the site returns to how it was within minutes. Nothing is closed until the site has been measured cold again, on the same basis as the first audit.
Will you guarantee a speed score?
No. Scores move from one run to the next, mobile and desktop differ, and third-party scripts set a floor no fix can pass. You get measured figures before and after, and a written view of what is within reach once we have seen your site.
What if we cannot deploy the files ourselves?
Choose either way. We send the finished files with instructions for whoever looks after your site, or you give us access limited to the web folder and we apply them after taking a backup. Your main hosting login is never needed.
Not sure what you need? Ask in writing.
Send the address; the reply says what is wasteful, what the fix involves and what it will measure afterwards.