AI can draft the homepage five ways. You pick on the real page.

Generating a page got cheap. Choosing did not.
An agent can put five finished homepages on the table in one sitting. Same URL. Each one is a complete server-rendered document, not a client-side skin. The scarce question is no longer "can we ship a page." It is "which of these is the page."
Page variants are that table. They sit next to in-page review mode, on the same ?review=1 overlay, and they do a different job. Review writes "change this sentence." Variants ask "keep this version."
How to declare the files, and how an agent should produce a batch in one pass, lives in the page variants guide. This piece is the scene: what the capability is for, and what it is not.
Look at the real homepage first
The pictures below are this site's homepage, not a blank demo page labelled Variant A.
On a local render, add exactly ?review=1. If the page has more than one version, the same review overlay grows a bar. On this homepage the names are default, aurora, cubegrid, hero-widget, topo, plus Select and Split.

Local Web Device after arc#4557 · this site's /en/ · 1440×900 · 2026-08-21. The bar is the review overlay, not a second toolbar.
Split is the part that changes the conversation. One click turns the page into a grid. Each pane is a finished render. The inner frames hide the overlay, so you compare pages, not toolbars.

Same environment, ?review=1&split=1. Five names on the bar, four panes (topo stays on the bar). Click a pane title to leave split.
That is the authoring loop the files were built for: produce a set, look at the set, pick. A/B testing in the traffic sense can come later, using the same directories. It is not the first thing the files are good for.
Two kinds of difference
Not every alternative is "a different page." Some alternatives only change values. Some change structure.
Values. Spacing, accent, type scale, a background colour. The blocks stay the ones readers already know. Missing pieces fall back to the default page. This is still a variant. It is not "switch to a built-in theme named dark." Different versions apply different concrete values.
Structure. A different hero, a different block order, a widget the default page does not have. The homepage pictures above are this kind: quiet default, aurora wash, topographic field, cube grid. Same headline. Different finished surface.

?review=1&variant=aurora on /en/. Still the real homepage, not a mock.
A third kind, a single block or a single sentence, is not in this wave. Whole-page split is the way to compare until that is no longer enough.
Traffic A/B is a later use of the same files. A visitor cookie can land someone in a bucket; search engines always get the default. Metrics and "is this significant" are out of scope. The new loop is earlier: an agent drafts the batch, a person selects, and only then, if you want, you open the split to visitors.
Next to review, not instead of it
In-page review is the other half of looking at a finished page. You mark a span, leave a page note, run checks. Variants do not replace that overlay, and they do not get a second one. Pages with only one version keep the review overlay as it already was.
On a page that did declare versions, the overlay grows three extra moves:
- The name buttons reload the same address as that version, for looking the way a reader would.
- Split puts the versions side by side, for choosing.
- Select writes which name won. If a page note is sitting in the overlay, that text can go with the selection as the reason. Empty is allowed. You can pick without writing a speech.
Visitor assignment is optional and later. Extra versions can exist only for the person and the agent who are choosing, while every public visitor still gets default, until you decide to open a traffic split.
The files are the interface. An agent that can put another finished homepage next to the current one has already made a version. It does not need a new API. The human job is to look, and to pick.
That is the whole point of putting several finished pages on one URL: generation is cheap; judgment is still scarce.