inside the machine: how AI assistants choose what to quote

When an AI assistant answers a buying question, it is not consulting an opinion. It is running a pipeline, and each stage of that pipeline is a competition your content either wins or loses. Understanding the pipeline is the difference between optimizing by superstition and optimizing by mechanism.

The assistant typically rewrites and expands the user's question into several search queries behind the scenes, a process often called query fan-out. Someone asks one question; the engine may search for definitions, comparisons, prices and reviews at once. Consequence: you are not competing for one query but for the cloud of sub-questions around it, which is why a cluster of related pages outperforms one page trying to say everything.

Retrieved pages are split into chunks, and each chunk is scored for how well its meaning matches the sub-question, using the same family of embedding techniques that power modern search. The engine does not care that your page was excellent overall. It cares whether some specific passage, often just a few sentences, cleanly answers the thing being asked. A brilliant answer buried in paragraph nine of an unrelated section is, to this pipeline, barely there.

The generating model reads the top-ranked passages and composes the answer, citing the passages it leaned on. Passages that are self-contained, factually dense and unambiguous are easiest to lift and attribute, which is consistent with the KDD 2024 GEO study's finding that clearer, better-sourced prose earns measurably more visibility. Ambiguity is not just bad style; it is a retrieval penalty.

Give every section one job. Open each section with its conclusion. Make each passage make sense if read alone, with the subject named rather than pronouned. Put the number, the source and the claim in the same breath. None of this is a trick; it is formatting truth so a machine under time pressure can safely repeat it.

Every major platform runs its own variant of this pipeline and none publishes the details, so this is the well-evidenced general shape rather than any one engine's blueprint. The practical bet is robust anyway: writing that wins under this model is also clearer for humans, so the downside of following it is roughly zero.