AI research
Two questions about AI, kept apart. How the answer engines themselves behave when you ask repeatedly, which is method work. And which AI vendors people are actually reaching for, which moves faster than any index tracking them. Each study below registers its question before any data is pulled, publishes every row, and is re-run on a schedule.
Google’s AI Overview does not cite the pages Google ranks
The question: Is answer-engine work a discipline of its own, or a second name for search ranking — does the generative answer cite the pages that already rank beneath it?
What it found: It does not. Across 8 high-demand questions, an AI Overview appeared on every one, and its citations overlapped the organic results on the same page by an average of 35%. Two-thirds of what it cites does not rank beneath it. The split runs by question shape: a named dish reached 83%, while a shopping task shared no domain at all.
Most answers are not researched. They are remembered.
The question: Before a content budget is justified on AI visibility, does an assistant actually read anything when answering the questions a brand cares about?
What it found: Not usually. Retrieval leaves a measurable signature — about 8,174 input tokens when pages are fetched against roughly 320 when they are not — and across eight questions carrying 2,404,500 monthly searches the largest assistant left none of it. The rule is not fact against opinion: it tracks whether the answer moves. Prices move, recalls move, recipes do not. Two earlier rules were written, tested and discarded.
Chinese AI demand is not falling. It is leaving DeepSeek.
The question: China’s AI index fell 41.7% from its July peak. Does a buyer holding an unsigned AI contract gain anything by waiting?
What it found: No. Western AI demand grew in all four markets measured, with no break at the July 1 boundary. Inside the Chinese arm, Kimi grew 309–831%, GLM 177–529% and Qwen 22–173% while DeepSeek — which is 78–90% of that arm — fell from a peak eighteen months older than the index’s. The category total was reporting one name’s decline.
What this category covers
How answer engines behave when the same question is put to them repeatedly — whether they retrieve, what they cite, and which vendors people search for.
What it does not
Endpoint behaviour on a named day, usually on one operator, in United States English. It does not see inside the decision to search, and the values move: one measured figure here doubled in two days.
When each study is read again
- Google’s AI Overview does not cite the pages Google ranks — When Google changes the surface. Last read 2026-10-29.
- Most answers are not researched. They are remembered. — When an operator changes an endpoint. Last read 2026-10-28.
- Chinese AI demand is not falling. It is leaving DeepSeek. — When the window extends. Last read 2026-10-28.
A study is corrected in place and dated when a re-read moves it. The correction is published; it is not applied quietly.
The six decisions behind these studies
These studies measure what AI assistants answer. The AI visibility Decision Brief takes the six decisions the seat funding that work is already arguing about and answers each one, with what was predicted before anyone looked.
See the six decisions$495 standard, $995 custom. Pre-made and delivered on payment. If the decision turns out not to be decidable from evidence, it is refunded.
Every study here is re-run, and the next reading will differ from the last.
Sound Decisions carries each re-run and each correction, with the registered question written before the data is pulled.
Subscribe to Sound Decisions