Performance is usually scheduled as a phase: design, build, then "optimise". By the time that phase arrives the expensive decisions have already been made, and what is left is compression. The result is a team squeezing kilobytes out of images while a hero section loads three typefaces nobody chose deliberately.
The decisions that cost the most
Almost everything that makes a page slow was decided before a line of code existed:
- How many typefaces, in how many weights, and whether any of them is needed above the fold.
- Whether the first screen is an image, a video, a canvas, or type.
- How much of the page needs to be interactive, and therefore how much JavaScript has to reach the browser.
- Whether the layout can be drawn before the data arrives, or only after.
- How many third-party tools the marketing team expects to add later.
Every one of those is a design decision. An engineer can execute any of them well and still ship a slow page, because the budget was spent before they were handed the file.
A budget is a design constraint
Treat transfer size the way you treat a column width: a number agreed at the start that the design has to work inside. Not a score to chase afterwards, and not a target that appears in a retrospective. A budget changes the conversation from "can we make this faster" to "what do we want to spend this on", which is a design question and has a design answer.
A performance budget is not a constraint on the design. It is part of the design, in the same way a page size is part of a book.
It also makes trade-offs visible while they are still cheap. A second display typeface is a real choice with a real cost; so is a hero video. Both can be worth it. Neither should be discovered in the last week.
What it looks like in practice
On this site the budget shaped the design system directly. Three typefaces, self-hosted and subset so a Turkish page loads no extra file. No animation library, because staggered reveals and a drawn trace are CSS plus one small observer. Every generated graphic is deterministic SVG rendered on the server, which is also why there is no stock photography anywhere.
None of those were compromises made to hit a number. They are the design, and the number is how we knew the design was finished.
The measurement that matters
Lab scores are a development tool. What a visitor experiences is field data: p75 LCP, INP and CLS on the devices people actually hold. A build can score well in a lab and still be slow in a market with older phones and worse networks, which is exactly the market most businesses are trying to reach.
That is also why performance and measurement belong to the same conversation: without field data you are optimising for a laboratory, and the laboratory is not buying anything.