An agent writes a site from one description. That part is genuinely one prompt. Then what?
We can count it. Across 2,400 published projects, the median site takes three versions, and the split is the interesting part: one before it goes public, two after. The first of those two lands about five minutes after the first version.
So the honest picture is not "one prompt and done". It is one prompt, a look at the result, and then a short run of small revisions.
The median site takes three versions
| Versions | |
|---|---|
| 25th percentile | 1 |
| Median | 3 |
| 75th percentile | 6 |
| 90th percentile | 9 |
| Mean | 4.51 |
The mean sits above the median because of a long tail: the busiest project in the set reached 49 versions. A handful of people keep a page alive for months and treat every change as a version, and they pull the average around without saying much about the typical case.
The distribution across the whole set:
| Versions | Share of projects |
|---|---|
| Exactly 1 | 27.8% |
| 2 to 3 | 27.8% |
| 4 to 6 | 22.0% |
| 7 to 10 | 14.6% |
| 11 or more | 7.8% |
More than a quarter never change again
27.8% of projects have exactly one version. Published once, never touched.
That is not a failure mode. A page for an event that has now happened, a demo sent to one client, a one-off explainer: none of them wants a second version, and the person who published it got exactly what they came for.
It does mean any average you read about "iteration" is describing two different populations at once. Over a quarter of sites are finished on the first try, and the rest have a median of four or five versions between them.
The work happens after publishing, not before
This is the number that changed how we think about the product. The median project has one version before its first publish and two after it.
Which is the opposite of how most tools are built. A builder assumes you perfect the page and then ship it, so publishing is the finish line and every setting sits before it. What people actually do is publish something reasonable and then improve it in public, at a real address, looking at the real page.
The reason is not carelessness. A draft in a preview pane and the same page on a phone at an actual URL are different objects, and the second one is where you notice that the headline takes four lines and the button is below the fold.
The first revision comes in about five minutes
Median gap between version 1 and version 2: 5.2 minutes. Between later versions: 0.4 days.
Two different rhythms, and the shift between them is the whole story of a small site.
The five minutes is one sitting. You look at what the agent made, you see the thing that is wrong, you say so, it publishes again. Nobody schedules that; it is a continuation of the same conversation.
The 0.4 days is life. The price changed, a new piece of work is worth showing, somebody pointed out a typo. Those revisions arrive in a different mood, days apart, and often after the page has done its job once already.
Any tool that makes the first gap expensive is charging you at the exact moment you are least willing to pay. Five minutes after publishing, the thing you want is a sentence, not a deploy pipeline.
Going back is rare, and it happens fast
11.0% of projects rolled back at least once. One in nine.
When it happens it is shallow and quick: the median rollback goes one version back, and arrives 23 minutes after the version it undoes.
Which reads as a specific kind of event. Not an audit weeks later that finds a mistake, but the same session, minutes after a change, when somebody looks at what they just asked for and decides the previous one was better. That is a reason to keep old versions addressable rather than to keep a changelog.
The other side is that eight projects in nine never go back at all. Rollback is insurance, and most people never claim on it.
These are 2,400 projects published through one host, mostly small sites written by an agent. It is not a measurement of how websites in general get revised.
It also says nothing about why. We can see that a version followed another five minutes later, and we cannot see what the person disliked.
What we take from it
Three things, and they are all about where to put the effort.
Publishing is not the finish line. The median site does two thirds of its versions after it is already public, so the address has to exist early and stay the same through every revision.
The five-minute revision is the one that matters most. It is the most common change a site ever gets, and it should cost one sentence.
Keeping old versions beats writing down what changed. One project in nine goes back, and when it does it wants the previous page live again within the hour, not a description of it.
Publish first, revise in public
The address stays the same through every version, and the old ones stay one request away.