What is query fan-out, stated plainly?
Query fan-out is the step where an AI search system takes one typed question and works on several derived questions instead. Google describes this decomposition in its own documentation of AI Mode, and the behaviour is visible to anyone who reads an AI Overview closely: the answer contains claims the original query never asked for, drawn from sources that answer a narrower question well. The practical consequence for a publisher is that the unit of competition stops being the query and becomes the set.
For ftes.ai this changes what a piece has to contain. A guide that answers only the headline question is a partial source for every assembled answer, and a partial source is a source the system can replace with a more complete one.
The sub-question set is predictable, and that is the whole method
Across the ftes.ai library the sub-questions a buyer's query fans into fall into seven recurring shapes: what it is, how it works, what it costs, what the risk is, how it compares, when to do it, and who is accountable. The shapes are not a discovery about Google; they are the shape of a purchase decision, which is why a decomposition step that models user intent lands on them. Writing the seven down for a given query takes about ten minutes and requires no tool.
The method is therefore: write the primary query, write its seven sub-questions in the reader's own phrasing, then read the page and mark each sub-question answered, partly answered, or absent.
- What is it — a definition sentence a stranger could lift
- How does it work — the named mechanism, in order
- What does it cost — a number, its unit, and where it comes from
- What is the risk — the failure mode and the gate that catches it
- How does it compare — the nearest alternative, with a verdict
- When should I do it — the trigger, not the season
- Who is accountable — the human who signs it off
How is each sub-question scored?
Answered means a reader who saw only that section, with no other part of the page, would have the answer. Partly answered means the page gestures at it without stating it — a benefit instead of a mechanism, a range instead of a number, an implication instead of a sentence. Absent means the page does not address it. The scoring is done by reading, by a person, because the test is whether a passage stands alone, and standing alone is a property a human reader can judge faster than a script can.
The score that matters is the count of absent and partly answered sub-questions. That count is the work list. A piece in the ftes.ai library ships when the count reaches zero for its primary query.
A worked example: what three of seven looks like
Take the primary query "what does a Digital FTE cost". Its seven sub-questions are: what a Digital FTE is, how it is billed, what the starting price is, what happens when the budget runs out, how it compares to a human hire, when metered billing beats a seat licence, and who controls the spend. A draft that opens with a price and a comparison table answers three of the seven. It reads complete to its author and is a partial source to a retriever.
The fix is not more words. It is four short sections: one naming token metering as the mechanism, one stating that a wallet at zero produces a hard stop rather than a surprise invoice, one giving the trigger for choosing metered over per-seat, and one naming the person who sets the budget. The draft grows by a few hundred words and moves from three of seven to seven of seven. The numbers in this example are illustrative of the method, not a measurement.
Why does one page have to carry the whole set?
Because assembly happens per sub-question, and a source is selected per sub-question. Spreading the seven answers across seven pages means competing seven separate times, each time against a page that covers more. Carrying all seven on one page means one piece is eligible for every slot in the assembled answer, and the internal links then work as reinforcement rather than as a prerequisite. This is the same discipline that makes a page useful to a human who arrived with one question and left with the decision made.
There is a limit. A page that carries sub-questions from a different purchase decision is not more complete, it is unfocused, and it will lose the sub-questions it should have won.
What this method cannot show
Coverage scoring cannot prove that a page will be cited, and ftes.ai does not claim otherwise. Retrieval depends on the index, the system's own source preferences, competing pages, and the phrasing of the query on the day. What the method does show is whether a page is eligible: a page missing four sub-questions cannot be selected for them, whatever else is true. Eligibility is the part a publisher controls, which is why it is the part worth a repeatable method.
The honest framing is the one the Search Visibility FTE uses in its own workspace: coverage is measured, citation is observed, and the two are reported separately rather than blended into a single confidence number.
How often should coverage be re-scored?
Re-score a page when its sub-question set changes, not on a calendar. The set changes when the product changes, when a new alternative appears that readers now compare against, or when the pricing model moves. A page about metered billing acquires a new sub-question the day a competitor ships a seat plan, because readers start asking which they should choose. A quarterly sweep over the pieces that carry the most commercial weight is enough for most libraries, with an immediate re-score on any page the product change touched.
The sweep is cheap: seven questions, one reading, a count. It is the cheapest reliable signal a content library has.