median page, on the lighter image variant
LAB · MEASURED ON SEPTEMBER 13, 2026
Anyone selling performance
should publish their own.
We looked at eleven of the best-known digital marketing agencies in Brazil and beyond. All of them sell speed and experience. None of them publishes the numbers for their own site. These are ours — updated with every change, including the ones that make them worse.
the same page when the screen asks for the larger images
the lightest page — the NØVA GRID demo, the most complex in the showroom
third-party scripts: the only one is the Google font. No ad pixel, no tag manager, no chat widget.
Weight per page
Bytes actually transferred, already compressed, measured against the production build. The light bar is what loads when the browser picks the smaller images; the dark one, when it asks for the larger. High-density screens sit near the dark end.
What the measurement shows and does not flatter us: the Work page gathers everything the studio has produced and is the heaviest on the site in the larger-image variant. It is queued to be split into on-demand loading. The pages with no photography — NØVA GRID, Plans, Brief — sit around 280 KB, which is the cost of the code.
Your session, right now
Measured by your own browser while you read this page. Nothing is sent: the number exists only here.
Interaction latency only appears once you interact with the page — hover, scroll, click. A dash means your browser does not expose that measurement.
Image library
the lossless originals, seven PNGs everything comes from
what the site serves today: 203 files in AVIF, WebP and JPEG, at the same resolution
Not one pixel was thrown away: the resolution is still 1536 × 1024. The quality of each file was calibrated individually, down to the threshold where the difference stops being visible.How that was done
Showroom studies
Every direction in the showroom answers a question. Here is each question — the answer is on the study page.
NOIR
Can a cybersecurity brand convey trust without a padlock, a shield and a photo of a hooded hacker?
AURAHow does a health network feel welcoming without falling into the cliché of the white consulting room and the generic smile?
MONOLITHHow does an infrastructure company show scale without a helicopter shot — and without coming across as cold?
VAULTCan wealth management inspire trust without a rising chart, a handshake and the word "exclusive"?
SIGNATUREWhat does a high-end website do differently from an expensive one?
LEXIUMCan a law firm look contemporary without losing its authority — and without marble, scales or a gavel?
PULSEHow does an AI company show usefulness instead of science fiction — and instead of promising?
ATELIERHow does an architecture studio let the work speak before the words — without turning into a photo catalogue?
FORGECan heavy industry get a website that does not look like a PDF from 2008 — and that actually serves to sell?
RAWHow far can a brand go before it turns into noise — and what separates provocation from noise?
Logbook
The record of what we decided and why. Not a tips blog: what actually happened on this project.
We published our own site's metrics
Every agency talks about performance using the client's site. We decided to measure ours and leave the number in plain sight, including when it is not flattering.
17 MB of images became 1.2 MB — and what that means in sales
The seven images in our portfolio weighed 17 MB. The fix was not better compression: it was changing format and measuring fidelity instead of guessing at quality.
Seven demos in our portfolio were broken
Seven of the showroom's ten directions returned a 404 error. The reason fits in one line of code — and it is a trap any modern React project can have.
How we measure
The script scripts/measure-site.mjs is in the site’s repository. It boots the production build, fetches every page and adds up what the browser has to download: the document, the CSS, every JavaScript file and the image each <picture> would choose. Bytes are counted compressed, because that is what crosses the network.
The first version of that script was wrong by a factor of five: it measured the uncompressed files. We fixed it before publishing, and the episode became an entry in the logbook. If a number here looks too good, be suspicious — and write to us.