Why did so many programmatic sites lose rankings?

The lesson usually drawn from the March 2026 core update is that programmatic pages are dangerous, and that reading is both comforting and wrong. What was penalised was a specific property: sets of pages where the only thing distinguishing one from another was a substituted noun. Ranking losses in the 60–90% range landed on page farms built by templating a city, an industry or a product name into otherwise identical prose. Volume was correlated with the failure because it is easy to produce that pattern at volume, not because volume caused it.

The useful version of the lesson is a test rather than a limit, and it can be applied before anything publishes: name, in writing, what appears on this page that could not appear unchanged on its nearest sibling. A page about acquisition in Texas that cannot name the licensing authority, the local market structure, the terminology used there and at least two figures that would be wrong in Florida is not a Texas page. It is a template with Texas in it, and no amount of editing the prose will change what it is.

That test is worth automating rather than promising, because a discipline enforced by a person reading pages stops being enforced somewhere around page sixty. The mechanical version is unglamorous: mask out every identity token — the place, the industry, the subject name — and compare what remains. Two pages that are near-identical with their names removed were always the same page. This is cheap to compute and impossible to argue with, which is the combination that makes a rule survive a deadline.

The second half of the shift is about citation rather than ranking, and it changes what the pages are for. Fewer than one in ten sources an AI engine cites appear in the organic top ten for the same query, which means visibility and ranking have partly decoupled. Generative systems suppress redundancy: a page restating a consensus already in the training data supplies no reason to cite anything. Original measurement, specific local substance and figures that exist nowhere else are what survive that filter — and they happen to be exactly what the differentiation test demands anyway.

So the conclusion is not to publish less. It is that the data floor comes first and the page follows, rather than the other way round. Where genuinely page-specific evidence exists, the page is worth building at whatever scale the evidence supports. Where it does not, no quantity of well-written prose creates it, and the page has nothing under it but confidence.