When an agent publishes a website, what is actually in it?
We looked at 2,400 published sites. 70.5% are a single file. The median one weighs 95.8 KB. 45.1% are not public. And while landing pages are the biggest category at 30.8%, the second biggest is not what anyone predicted: reports and dashboards, at 17.7%.
Put together, those numbers describe a kind of web page that barely existed five years ago.
Seven in ten are one file
| Files | Share of sites |
|---|---|
| 1 | 70.5% |
| 2 to 3 | 14.2% |
| 4 to 10 | 12.0% |
| 11 or more | 3.3% |
One index.html with its styles and scripts inline. No folder structure, no assets
directory, nothing to wire up.
That is what a model produces when you ask it for a page and do not tell it how to organise the result. It is also, not coincidentally, the easiest possible thing to host: one file has no relative paths to break, which is the failure that ruins most attempts to move a generated site somewhere.
The 3.3% with eleven files or more are a different animal, usually a small multi-page site that grew section by section.
The median site weighs 95.8 KB
Median 95.8 KB, mean 227.6 KB, and 51.2% come in under 100 KB.
For scale, HTTP Archive's 2025 page weight chapter puts the median desktop home page at 2,862 KB, up 7.3% on the year, of which 697 KB is JavaScript and 1,059 KB is images.
The two figures do not measure quite the same thing, and it is worth being precise about it. Theirs is bytes transferred to load one home page, including everything pulled from third parties. Ours is the total size of the files that make up the published site. Close enough in kind to compare orders of magnitude, not close enough to treat as one metric.
Even allowing for that, the gap is about thirty times. The reason is subtraction rather than optimisation: no tag manager, no cookie banner, no chat widget, no font stack, no framework. Nobody minified anything. The page is small because nothing was added to it.
Nearly half are not public
45.1% of published projects are private, and 12.2% carry a password.
This is the number we underestimated most. Publishing, for a lot of people, does not mean putting something on the open web. It means turning a thing into a link they can send to one person.
A draft for a client to approve. A report for a team. A page shared with three people who need to look at the same thing at the same time. The address is the point, and being findable is not.
It also explains why the usual measures of a website miss this category entirely. Nothing in this half of the set has SEO, or traffic, or an audience. It has a recipient.
What these sites are for
| Category | Share |
|---|---|
| Landing page | 30.8% |
| Report or dashboard | 17.7% |
| Documentation | 12.6% |
| Portfolio or CV | 7.8% |
| Form or survey | 6.6% |
| Presentation | 6.5% |
| Calculator or tool | 6.3% |
| Game or demo | 3.8% |
| Other | 7.9% |
Landing pages leading is unsurprising. Reports and dashboards sitting second, ahead of documentation and portfolios combined, is the finding.
Think about what a report as a web page replaces. Somebody had numbers, asked an agent to turn them into something readable, and got a page instead of a spreadsheet or a slide deck. It renders the same for everyone, opens on a phone, and needs no software to read. That is a use for the web that a website builder never targeted, because it was never a website in the first place.
Presentations at 6.5% are the same instinct pointed at a different artefact, and calculators at 6.3% are the version where the page has to do arithmetic.
The CDN detail, without the conclusion
57.5% of these sites load something from an external CDN: a font, a stylesheet, a small library.
It is tempting to draw a line from there to why generated sites break when people move them between hosts. We are not going to, because this data does not support it. A share of sites using a CDN is a fact about how they were written, not evidence about what breaks them. That would need a different measurement, of failures rather than of markup.
What it does mean practically is that "one self-contained file" is not the same as "no network dependencies". Most of these single files still reach out for something on load.
These are 2,400 sites published through one host by people working with an agent, so the set describes that population and not the web.
The categories are a classification of content, not a statement of intent. A page we read as a report may have been a landing page to whoever made it.
Why the shape matters
A single 96 KB file that half the time is not even public is a different object from what hosting was built around, and the difference has consequences.
It needs no build. Nothing here compiles, so any step between the file and the address is overhead somebody else invented.
It needs an address more than it needs an audience. For nearly half the set, the whole product is a link that works and keeps working.
And it changes more often than it is visited. Which is a different article, and we counted that too: the median site takes three versions, two of them after it is already live.
Publish the file, get the address
One file or forty, public or private. Your agent publishes it and the address stays.