The Same Audit Says a Day of Drift Is Still Smaller Than an Engine Change

The same September 2026 paper separates a change of day from a change of engine. A 5-to-6 June comparison inside one engine found mean URL-set turnover of 67.0 percent, with a 95 percent confidence interval of 61.1 to 72.2 percent. Two-thirds of the URL set moving in a day is a large number. The abstract's next sentence is the comparison that matters for measurement: a same-engine citation set on the adjacent day was still about 42 times more similar than a same-day set taken from two different engines.

The author treats that gap as the point of the audit. An engine-free score can estimate page quality or how well a page fits a query. End-to-end visibility, the abstract says, also depends on exposure and selection that belong to a particular engine. The paper proposes reporting four quantities separately: page fit, observed exposure, conditional selection, and final visibility.

What the audit does not claim

The benchmark is 15 commercial prompts on four named surfaces, not a sample of the open web. The author writes that the results do not invalidate engine-free page scoring. They identify what that score is estimating. A score that never observes an engine, the abstract says, should not be read as a stable probability of being cited by that engine unless an exposure model is stated.

Source: arXiv 2609.22655.