React 19 puts the deleted head tags back before page.content() serializes them

React owns the nodes it hoisted. Calling `remove()` on them inside `page.evaluate()` unmounts nothing. It desynchronises the live DOM from the tree React believes it rendered, and the next commit restores what was deleted ^[10].

The duplicates get there in the first place because a client-side route transition briefly mounts one route's component before resolving to another, both render an `SeoHead`, both sets of tags get hoisted, and the unmounted one is not always cleaned up before the page is serialized ^[9]. The window the reinsertion lands in is structural: `page.evaluate()` and `page.content()` are separate CDP round trips, the page keeps running between them, and a lazy chunk landing or a queued state update or an effect firing is enough for a reconcile ^[11].

For in-page deletion to hold, the page would have to be quiescent at the instant of serialization, and the prerender script has no way to establish that. It awaits `domcontentloaded`, then a `waitForReactToMount` helper, then reads content ^[8]. A mounted component only confirms it exists, not that it has settled; the script never checks for the second condition. As the author of the writeup puts it, you are in a fight with the reconciler and the reconciler runs last ^[17].

The build's dedupe check runs against the live DOM rather than the file that ships ^[12], which is how it keeps a perfect pass rate while doing nothing.

The spec detail is what makes the rewrite reliable. `document.title` returns the child text content of the first `<title>` element in the document, which is also the one the browser shows and the one Google reads ^[14], so it answers which of three titles actually counts. `cleanHead` then keeps that first title and drops the rest, drops `link rel="canonical"` tags with no `href` and keeps exactly one, and for `meta[name]` or `meta[property]` drops empty-content placeholders where a populated tag exists for the same key ^[15]. The reason this cannot lose the same race twice: once the head has been serialized, it is a string, and there is no reconciler left to write back into it ^[16].

The volume is small and worth stating anyway. Seven head tags per page where three carry the meaning leaves four redundant, and the site has 23 pages, so a build in which every page is prerendered ships 92 tags nobody wanted ^[1].

There's a limit to what this evidence explains. Of 23 pages, 22 sat in Google Search Console as "Discovered - currently not indexed" ^[1], and the author's conclusion is that this is mostly a new domain that nobody links to ^[2]. The duplicate metadata was found on the way to that answer, not shown to have caused it.

This reaches your build only when several conditions in the source line up: React 19's hoisting of `title`, `meta` and `link` from anywhere in the component tree ^[6], a router that transiently double-mounts route components so two `SeoHead` outputs coexist ^[9], and a prerender step that serializes a live page rather than a string ^[8]. Drop any one of those and the collision cannot happen; keep all three and the dedupe step reports success on every page.