Query fan-out and what it means for your content

This is the most useful mechanical fact in AI search, and it is the one least often explained to business owners, because it does not sell anything.

Once you understand it, a lot of odd behavior stops being odd. Why you appear for phrasings you never targeted. Why a competitor with a worse main page gets cited anyway. Why your carefully optimized service page loses to somebody's FAQ.

What does fan-out look like in practice?

Google's own description is short. Its generative features use retrieval and grounding from the Search index, plus query fan-out, which it defines as "a set of concurrent, related queries generated by the model." That is the whole mechanism, stated by the operator.

Here is what it looks like from the customer's side. Someone types a real sentence, not a keyword. Something like: is there anywhere near downtown that can do a party of twelve on a Friday and handle a gluten allergy.

The system does not search that string. It decomposes it. Restaurants downtown. Restaurants with large party seating. Friday availability. Gluten-free handling. Possibly recent reviews mentioning groups. Then it retrieves for each and writes one answer over the top.

Your restaurant might be retrieved for the group seating search and for nothing else. A competitor gets retrieved for the gluten question because their allergen page exists. A third place gets retrieved because two reviews mention a birthday party for fourteen.

The answer is a collage, and each fragment was won separately. That is why the old mental model of one page targeting one query has stopped describing what happens.

You can watch this yourself. Ask a compound question, look at the citations, and note that they rarely all come from businesses that rank for the whole sentence.

What does this change about keywords?

It changes what you are optimizing against. Not the phrase, the components.

The unit of competition is now the sub-question, and sub-questions are boring, specific and operational. Do you take walk-ins. Is there parking. Do you serve after nine. Do you handle commercial as well as residential. What does it cost roughly. How fast can someone come out. Do you speak Spanish.

Most businesses answer none of these anywhere in indexed text. They answer them all day on the phone.

That gap is the opportunity, and it is unglamorous work: writing down what you already tell customers, in the words customers use, on pages that can be found. No new discipline required.

It also changes measurement. If you track one keyword and one position, you are watching a door people stopped using, which is part of why a single AI visibility number fools you and why prompt-level sampling with real sentences beats keyword rank tracking here.

Should you chunk your content for the machine?

No. Google names it directly, and it names three of its neighbours in the same breath.

In the Mythbusting section of its guidance on generative AI features, Google lists as ineffective: special files such as llms.txt and similar markup, artificial content chunking, rewriting content specifically for AI systems, and seeking inauthentic mentions. Two of those four are exactly what "AI content restructuring" packages sell.

There is independent evidence pointing the same way, and it comes from the one peer-reviewed experiment in this area. The KDD 2024 paper "GEO: Generative Engine Optimization" tested nine content tactics against a 10,000 query benchmark, scored on a position-adjusted word count metric. The do-nothing baseline scored 19.3. Keyword stuffing scored 17.7, below the baseline, meaning it actively hurt. And writing in a more authoritative tone produced no significant improvement at all. The authors' reading is that these systems already shrug off changes of that kind.

Both of the failed tactics share a shape: they are things done to text to please a machine rather than a reader. The tactics that did work in the same experiment are the opposite. Adding direct quotations scored 27.2, adding verifiable statistics 25.2, and adding inline citations 24.6. Those are things a careful writer does anyway.

So the honest instruction under fan-out is not "chunk your content." It is: answer one thing per section, answer it in the first sentence, and support it with something checkable.

What is an AI content restructuring package worth?

Zero, and the alternative is cheaper than the quote.

The quote. An agency offers "AI content restructuring" across twelve pages at $4,800, which is $400 per page (derived). The scope is chunking and rewriting for machine readability.

What Google says about the scope. Both named as ineffective, in the same list, on the same page. Price the line item at zero.

Now do the work it should have been. Write down the twelve questions your staff answer most on the phone, verbatim. Check which are answered anywhere in your indexed text. In my experience the answer is two or three, so call it nine missing, which is 75%derived of your operational sub-questions absent from anything a retrieval system can reach.

Price that. Nine sections, each a clear heading and three or four honest sentences, is about forty five minutes apiece. Under seven hours of your own time. If you would rather buy it, the same $4,800 at a writer's $150 an hour is 32 hours (derived), or roughly three and a half hours per missing answer. That is a generous budget for writing down what you say on the phone.

The comparison that matters. One purchase reformats text that already exists. The other adds nine claims about your business that did not exist anywhere retrievable. Under fan-out, only one of those creates a new way to be found.

Where does the mechanism break down?

At the point where people turn a benchmark result into a forecast, which the GEO paper's own data forbids.

Read Table 2 of that paper carefully. When every source in the candidate set is optimised, the Cite Sources tactic moved the fifth-ranked source by +115.1% while moving the first-ranked source by −30.3%. Quotation Addition moved rank five by +99.7% and rank one by −22.9%. This is a redistribution finding, not a rising tide. If you are already the strongest source in a candidate set, the same tactic applied by everyone costs you.

Two more limits travel with it. The same tactic ranged from 19.1% to 35.8% in effectiveness depending on the subject category, so there is no universal number. And the experiment was a simulation: cleaned text from five fixed Google results, fed to an older model, sampled repeatedly. A validation on Perplexity with 200 queries via file upload found the best methods gained 22% on word count and 37% on subjective impression, which is corroboration of a mechanism rather than a live measurement.

And nothing in the benchmark is a local commercial query. The paper has not been validated for local intent at all. So take the direction, which is that verifiable, quotable, cited writing gets used more, and leave the numbers where they were measured.

You may not combine those figures. The +115.1% rank-five lift and the 19.1% to 35.8% domain range are different metrics measured under different conditions, and neither is a percentage of anything you can count. Multiplying them, or applying either to your traffic, produces a forecast with no source under it.

Which of your pages actually win under fan-out?

Pages that answer one thing clearly and completely.

A page that answers a specific question in its first three sentences, then supports the answer, is easy to retrieve for that sub-question. A page that covers eight topics at medium depth is a weaker match for all eight.

Self-contained paragraphs matter for the same reason. A paragraph that only makes sense after reading the two above it is a bad candidate for extraction. One that stands alone can be lifted into an answer intact. Note the difference between that and chunking. Writing paragraphs that stand on their own is composition. Splitting a page into machine-labelled fragments is the thing Google names. The first is how good writing has always worked.

Your business also needs to be recognizably one business across all these separate retrievals. If three sources describe you slightly differently, the system has to reconcile them, and reconciliation sometimes fails. Entity consistency across the web is what makes fragments won separately resolve to one business.

Some sub-questions are about current state rather than permanent facts. Whether you are open, what is running this week, what changed. Your Business Profile is the natural home for that, and whether profile posts affect anything has a more interesting answer once you think of them as answers to time-sensitive sub-questions.

Reviews cover the sub-questions you would never write yourself, which is the strongest argument for a permanent asking habit. Just keep the mechanics clean, because there is a compliant way to get staff asking and several non-compliant ones that will cost you more than they return.

Does this mean writing more pages?

Usually fewer pages, better organized.

The instinct is to spin up a page per sub-question. That produces thin duplicates competing with each other, and it is the same mistake as the old location-page farms with a new justification.

Better: take your existing service pages and make each one genuinely answer its question, then add the operational specifics as clear sections with honest headings. A page that answers eleven real sub-questions in eleven clean sections is retrievable eleven ways.

The exception is when a sub-question is genuinely large enough to deserve its own page. Pricing, process, service areas and requirements often are. If you would spend six minutes explaining it on the phone, it can carry a page.

What to do this week

1. Write down the twelve questions your staff answer most on the phone. Verbatim, in customer language. 2. Check which of them are answered anywhere on your indexed pages. Count the gap and run the arithmetic above. 3. Add the missing ones as clearly headed sections on the most relevant existing page, answering in the first sentence. 4. Rewrite the first three sentences of your top three pages so each answers its own question with no warmup. 5. Where you make a factual claim, add the number and where it came from. That is the one tactic the peer-reviewed experiment supports. 6. Ask one compound question about your category in an assistant, list the citations, and see which sub-question each was probably retrieved for.

Then decide what you can see downstream, since tracking referral traffic from AI assistants is harder than it sounds and worth setting up before you make changes rather than after.

Want to know which sub-questions you are missing?

Send me one service page URL at eric@seod.com and I will list the sub-questions a customer would ask around that service, mark which ones the page currently answers, and flag the two most valuable gaps. No charge, no pitch, and it is usually a short list of things you already say out loud daily.

I offer this because the gap between what a business tells customers on the phone and what it has ever written down is the cheapest content win available, and nobody sells it because there is no product attached.

Otherwise, there is more on AI search and AEO here.