The 35% Lift Was The Tide: What A Bundled Commit Hid About Which Fix Worked

A developer writing on dev.to shipped two changes to a travel site in a single commit: a roughly 4,600-line prerender fix across nine files, and a fifteen-line correction to a URL lookup added almost as an afterthought ^[3][5]. Two weeks of search data later, he reports the afterthought did most of the work, and the only reason he can say so is that he went looking for a control ^[6][8]. The headline work was real. 220 country pages had never been prerendered, so every crawler that asked for one got the raw single-page-app shell: identical bytes for all 220, canonical pointing at the homepage ^[1]. Google had filed them as duplicates of each other and of the front page ^[4]. Undoing that meant extracting a 2,800-line data structure from a React component that no build script could import because it carried icon references, shrinking the component from 4,343 lines to 1,417 ^[2], a removal of 2,926 lines from one file ^[6]. Then the measurement problem. The pages he fixed went up, and he nearly stopped there ^[25]. But the entire site was 35% ahead that week, including sections he had shipped nothing to ^[7]. So he used the blog, untouched in the same window, as a baseline: it ran 19% ahead ^[8]. Comparing eleven days before the deploy to nine days after ^[9], a twenty-day window ^[5], the 220 newly prerendered country pages came in at +41% ^[10] and the hub page linking to them at +171% ^[11], with the hub's average position moving from the high teens into the low teens ^[12]. Against the control, that is about 22 points for the pages he spent two days on and about 152 points for the page he changed by accident ^[1][2]. The detail that settles it: 142 of the 220 country pages were already earning impressions before the fix, and only eight appeared for the first time afterward ^[13][14]. Of the 78 pages that had no impressions beforehand, 70 still had none ^[3]. The duplicate-hell rescue mostly rescued pages that were not in duplicate hell. The fifteen lines fixed something the toolchain could not see. The hub built its country links from the database, joining each row's region and slug into a path ^[15]. The router resolves URLs against a static data file instead ^[16], and ten countries disagreed across the two ^[17]. Some disagreed on spelling: the database said `united-kingdom` and `saint-lucia`, the router knew `uk` and `st-lucia` ^[18]. Some disagreed on geography: the database filed Georgia, Armenia and Azerbaijan under Asia, the static file under Europe, both defensible, each populated by different people at different times ^[19]. So the hub emitted links to ten URLs that did not exist ^[20], and none of them 404'd. The route pattern `/world/[continent]/[country]` matched, the page rendered with no country to render, fell through to generic world-map copy, and returned status 200 with real HTML, a title, and a canonical tag pointing at itself ^[21]. Ten near-identical self-canonical pages, each endorsed by an inbound link from the most authoritative page in the section ^[22]. The link checker only asks whether a URL errors, the typechecker only asks whether a string is valid, the build only asks whether anything threw, and the sitemap never listed these URLs at all ^[23]. The author calls this identifier drift rather than fact drift ^[24]. The operational point is about attribution, not SEO. Line count is 300 times larger on the side that moved less ^[4]; a bundled commit plus a rising tide would have taught the opposite lesson and sent the next two days into more prerendering.