Because a complex question is almost never a single question — it's a group of sub-questions with different weights, and no answer engine covers all of them. A study from the Georgia Institute of Technology and Salesforce AI Research (NAACL 2025) decomposed 200 complex questions into around 4,000 sub-questions and measured how many showed up in the answers of three commercial engines: You.com, Perplexity AI, and Bing Chat. All three do prioritize core sub-questions over secondary ones — but they still miss around half of those core sub-questions.
It's the criterion that carries 12% weight in Nostos: Sub-Question Coverage. It doesn't measure whether you touched the topic — it measures whether you answered every part that matters.
It's the part of your question without which the answer stays incomplete — not an extra detail, but a necessary piece of the reasoning. The paper classifies every sub-question into three types:
That all three make the right distinction: they cover core sub-questions more than background or follow-up ones. But none of them consistently retrieves all the information needed for the core ones, and even when that information was already among the cited sources, it often didn't make it into the final answer. Two more findings from the same study: core sub-question coverage predicts which of two answers a human prefers with 82% accuracy — more than directly asking a model to compare the two answers, which reaches 71%. And when a RAG system prioritizes core sub-questions during retrieval, the resulting answer beats the baseline without that step up to 74% of the time.
Decompose your own topic before writing, not after. Four concrete steps:
Get your GEO score across eight academically-backed criteria, including your Sub-Question Coverage.
Run the analysis at nostosgeo.app