Inside the research
Short sections were the strongest intervention
In this experiment, self-contained sections under question headings beat the plain-paraphrase control on retrieval scores for 84–92% of the 49 questions across the four methods. The report also records a 20–29 percentage-point improvement over the control in how often the top passage contained the answer. These are results within the study’s retrieval setup, not observed citation gains in live AI search.
Naming entities and moving answers produced different effects
Explicit entity names produced smaller, more mixed gains in selecting the answer-bearing passage. Answer-first rewrites showed no statistically significant retrieval-score improvement over the control in this setup. Combining all three interventions performed similarly to short sections alone.
Restructure around questions, then test
The practical hypothesis is to give one question a clear section, include the context needed to understand that section independently, and compare the result with a control. The report suggests starting with conversational pages whose answers are spread across stories and examples. Its suggested section lengths are an experimental approach to test, rather than a universal rule for AI search.