The evidence bar for shipping.

Generative tools removed the cost of producing a page. What they did not remove is the cost of publishing one that should not exist. When a pipeline can generate a hundred plausible landing pages before lunch, the scarce discipline is no longer writing — it is refusal.

The rule: two independent signals

Nothing we publish programmatically ships on one signal. A candidate needs demand evidence from two independent sources — not two searches of the same index, not the same community read twice, but two mechanisms that could disagree with each other and did not. On top of that, the thing being described has to actually run: a real execution against the real source, checked for whether the returned data matches what the page promises.

Fail either test and the candidate is rejected, no matter how good the copy looks. There is no total score that lets a strong headline compensate for absent demand.

What it cost this week

We were building a batch of pages across our data catalog. One of them targeted property listings — a category where we have working infrastructure, a plausible audience, and every commercial reason to want a page.

The search-trend signal was there. The second source never arrived: four separate attempts to find real people discussing the problem returned nothing usable. Under the rule, that candidate does not ship — so it didn't. We replaced it with a different topic where both signals converged, and shipped one page fewer in that category than planned.

The pipeline also told us something we would not otherwise have learned: the reason the second source came back empty was partly a fault in our own tooling, not an absence of demand. A bar that is allowed to fail generates diagnostics. A bar that always passes generates only output.

Why the bar has to be mechanical

Judgment does not scale and does not stay honest under deadline. If the decision to publish is a feeling, then on a slow week the feeling becomes "close enough" — and the drift is invisible, because every individual page still looks reasonable. Ten pages later you have a category of content that answers nobody's question, and search engines have a name for that pattern.

Written as a rule, the same decision is auditable. Every rejection carries its reason. Every publication carries the two sources that justified it and the execution record that verified it. Six months from now, when something in the category is working or failing, the trail exists to explain why.

The general form of the problem

This is not really about landing pages. Any system that generates faster than it can judge needs an explicit standard that it is permitted to fall short of — and a habit of treating the shortfall as information rather than as an obstacle to route around.

The temptation of cheap production is to measure yourself by volume, because volume is the thing that suddenly got easy. We measure by what survived the bar. Some weeks that number is lower than planned. That is the number working correctly.

← All notes