Spoke
How an AI answer is assembled
It reads the question, fans it out into several searches, retrieves passages, writes an answer, and credits some of it. Knowing the order tells you which stage you failed at, which is the only way to fix anything.
Query fan-out, the eligibility conditions and the shared index are documented by Google in its guidance on AI features. The step between retrieving and citing is published by nobody, and this page says so instead of filling the gap.
A generated answer looks like one thing happening. It's a pipeline, and each stage fails differently.
How does an AI search engine decide what to put in an answer?
It works out what's actually being asked, fires off a set of related searches rather than one, pulls candidate passages from an index, writes a response from them, and credits some of it to sources. Google documents the fan-out step and documents what makes a page eligible: indexed, and allowed to show with a snippet.
What nobody documents is the step in between — how it picks among the passages it pulled. That's the question everyone wants answered and the one nobody outside those companies can answer honestly.
The stages
Interpretation. It works out what's being asked, including the parts nobody typed.
Fan-out. Google documents a query fan-out technique: several related searches generated at once to go get more information, with results fetched for each. This is the stage that breaks the idea of "ranking for a keyword." The system is researching a topic, not looking up a string of text.
Retrieval. Candidate passages come back. Your page is in that pool or it isn't, and that comes down to ordinary indexing and eligibility.
Synthesis. The model writes an answer from what it's holding. One sentence might draw on four sources.
Citation. Some sources get named, some don't, and how those two sets relate isn't published.
The three ways you lose
Here's the practical value of knowing the pipeline. Three completely different failures look identical from where you're standing.
Retrieval failure. You weren't in the pool. The causes are boring and technical — not indexed, snippet suppressed, not matching any of the searches that actually ran. This is the most common one and the most fixable.
Extraction failure. You got retrieved, but there was nothing in your page that worked as a standalone answer. That's what the AEO section is for, with the caveat recorded there that Google says no particular style is required.
Selection failure. You were retrieved, you were quotable, and it picked somebody else. Nobody can tell you why. Including whoever is selling you a fix for it.
Being straight about that third category is most of what separates a diagnosis from a pitch.
What fan-out changes for you
If the system is running several related searches, then covering the neighboring subtopics isn't padding. It's more surface for retrieval to catch.
A page that answers one narrow question beautifully may never get pulled in for the fan-out that actually ran.
More in Search fundamentals
How search works
Three stages, three distinct failure modes, and one diagnostic habit that saves most of the time people waste guessing. Everything else on this site assumes this model.
Semantic understanding
Search systems model meaning rather than matching strings, which is why naturally phrased content now ranks for queries that contain none of its words — and why keyword density stopped being a lever a long time ago.
Penalties
The tactics search engines act against have barely moved in two decades. AI didn't add a new category. It made the oldest one cheap.