Hi everyone, this is Neo.

This one starts with a SEJ piece that made me smile. The author, Greg Jarboe, opens by saying: “I was doing query fan-out by hand for two decades before anyone called it that. I just called it something else: Russian nesting dolls.”

His method was dead simple. When optimizing a press release, he’d hunt for a four-word phrase that naturally contained a three-word phrase. Say the core term was “airfare to Philadelphia” — he wouldn’t build around that three-word version. He’d find the four-word version, something like “cheap airfare to Philadelphia,” and write the whole release around the longer one.

The reasoning is just as simple: use the long version and you can be found by someone typing the short query or the long one. Use only the short version and anyone typing the longer phrase will never find you — because your page never mentions it.

That trick worked in 2003. And in 2026 it matters far more than it did back then — because it isn’t only humans searching anymore. AI searches too, and AI splits a single prompt into a pile of sub-queries before it does.

1. Query fan-out, plainly

Query fan-out is what happens when an AI search engine refuses to match your whole sentence against a page. Instead it decomposes the question into multiple related sub-queries, runs them in parallel, and synthesizes the results into one answer.

Google confirms this in its own documentation:

“Both AI Overviews and AI Mode may use a ‘query fan-out’ technique issuing multiple related searches across subtopics and data sources to develop a response.”

How many sub-queries? Estimates vary, but they cluster:

Source	Sub-queries per search
SEO consultant Aleyda Solis and multiple analyses	roughly 8–12 parallel sub-queries
Some AI Mode field tests	up to about 16 sub-searches in a single run
MJ Cachón’s branded-prompt dataset (August 2026)	189 prompts → 1,797 sub-queries, about 9.5 per prompt, averaging seven words each

In other words: one sentence typed by a user becomes a dozen short queries you never see. Whether your content gets cited depends on whether you can win those sub-queries — queries you have never researched, which are often shorter and more specific than what the user actually typed.

2. The most valuable finding inside Cachón’s dataset

The value of that dataset isn’t the volume. It’s the order.

She found that branded fan-out doesn’t behave randomly. It follows a pattern:

1. The first sub-query is plain, conversational language — the way a person would ask;
2. Then the site: operator starts appearing — the model is fencing off trusted ground and searching specific domains directly;
3. Further in, it starts pulling exact quoted phrases — to check whether a source actually says what the model thinks it says.

And step three intensifies along the chain: between the first search in a run and the last, quoted-phrase usage climbed 25-fold.

That single pattern carries two implications:

First, “being retrieved” and “being trusted” are two different gates. The early sub-queries decide whether you make the candidate pool. The later quoted searches decide whether the AI dares put you in the answer. The first gate is about coverage. The second is about whether your page contains a sentence that can be lifted verbatim.

Second, AI retrieval runs wide-to-narrow, while classic SEO ran narrow-to-wide. Jarboe’s nesting dolls expanded outward from a small phrase so that queries of any length would land on his page. Today’s models narrow inward until they hit a literal phrase. Two directions, one requirement landing on you: your content has to contain the same idea at multiple lengths. Skip one, and you vanish from that slice of the funnel.

3. Three numbers that bring this down to earth

Number one: 15% of queries are brand new every day.

Google wrote in its 2019 BERT announcement that 15% of the queries it sees on a given day are ones it has never encountered before. At Search Central Live NYC in March 2025, John Mueller brought it up again, sounding almost resigned:

“I would have thought at some point most of the searches would have been made; people just ask the same thing over and over again. But when we recalculate these metrics, it’s always around 15%.”

He expected large language models to push the number up. They didn’t. It just sits there.

Multiply 15% by the billions of searches Google handles daily and you get hundreds of millions of never-before-typed queries every single day. Most of them aren’t random — they come from something that just happened: breaking news, a new product name, a policy that just changed, a phrase a journalist coined on deadline that a thousand people typed into a search box within the hour.

Number two: AI-era queries are three times longer.

Google’s May 2026 AI Mode usage report says it plainly: the average AI Mode query is triple the length of a traditional search query, and traditional search hovered around three or four words for most of its history. The same report covers over 1 billion monthly active AI Mode users globally, follow-up queries growing more than 40% per month, and AI Overviews appearing on somewhere between 13% and 20% of queries depending on whose measurement you trust (some trackers put it closer to half).

Length isn’t a side effect of AI search. It’s the terrain now. And Cachón’s data plugs straight into it: one branded prompt fans out into sub-queries averaging seven words.

Number three: 73% of sub-queries are different every time.

One analysis of sub-query stability produced a bracing result: only 27% of fan-out sub-queries stay stable across repeated searches. 73% change.

That sounds like bad news and is actually good news: it kills off the strategy of sniping one specific sub-query, and it leaves full topical coverage as the only resilient approach.

4. So what should you actually change?

Jarboe offered three habits. Here they are translated into terms an independent site owner can act on — plus one of mine.

1. Find the nested phrase, not just the seed phrase.

Take a core term and write down two or three longer phrasings that naturally contain it. Then build your opening paragraph or your H2 around the longer version.

Where do you find those longer versions? Search Console, sorted for high impressions and low clicks. Those are almost always longer variants already knocking on your door. Finish the sentence they’re looking for, and the clicks come back.

2. Publish at the speed of the news, not the speed of the content calendar.

That 15% of brand-new queries is tightly bound to things that just happened. If your team has a same-day channel — a press release, a rapid-response page, a fast-turnaround slot on the company blog — that’s your best shot at owning a phrase before a competitor even knows it exists.

This is exactly where independent sites have an edge: your decision chain is short. A big brand needs two weeks of approvals. By the time they publish, most of the window has closed.

3. Write the answer as a standalone, liftable sentence.

Cachón’s data shows AI leaning harder on exact-phrase verification. If your answer only makes sense next to its surrounding context, it can’t be quoted. Rewrite it until it stands alone and still carries the information.

There’s a quantifiable benchmark here: an analysis covering 15,847 AI Overview results (December 2025) found that passages in the 134–167 word range achieved the highest citation rates. Not too long, not too short — roughly one paragraph that fully settles one question.

4. One more from me: apply the nesting to your product pages and heading structure.

The first three habits are about content. But what gets retrieved daily on an independent site is usually product pages. My approach is to put nested phrasing directly into H2s and spec subheadings:

• Use the long form in the H2 (“Customs clearance software for small and mid-size freight forwarders”);
• Include the short form naturally in the first paragraph (“freight forwarding software”);
• Use full expressions instead of abbreviations in spec tables — when AI quotes you, “warehouse management system (WMS)” is safer than “WMS” alone.

5. Don’t over-read this: it isn’t new magic

Three boundaries, or this piece just becomes more anxiety fodder.

First, Google’s official position is that optimizing for AI search is still SEO. There is no hidden AI-only layer. Anyone selling fan-out as a brand-new discipline is selling anxiety.

Second, don’t stuff variants to satisfy fan-out. Cramming every length variant onto a page is the same old keyword-stuffing habit in a new costume, and it gets judged as low quality in the AI era just as it did before.

Third, new-query windows are short. Those 15% new queries spike fast and collapse fast. They’re an opportunity, not a stable traffic channel. You can’t build a monthly report on them.

Neo’s take

What I love about Jarboe’s piece is that it captures a pattern in this industry: new concepts are usually just names for old manual work.

He’d been doing this for twenty years; back then it was called nesting dolls. Once it got named query fan-out, it became something you could teach, sell, outsource and productize. The naming created a market. It didn’t create the value — the value was a writing discipline that was already working two decades ago.

That’s a useful reflex for independent site owners: every time a new term shows up, ask whether the underlying action worked before the term existed. If it did, it’s a fundamental wearing a new coat. If it didn’t, then it’s genuinely new.

My second observation is more practical: fan-out raises the weight of breadth, but it doesn’t lower the weight of depth.

73% of sub-queries changing every time means you can’t predict every angle — you can only cover the topic surface. But quoted-phrase usage climbing 25-fold, and 134–167 word passages earning the highest citation rates, means what ultimately gets trusted is one paragraph that settles one thing properly.

Put those together and the conclusion is almost old-fashioned:

Cover wide enough to hit the sub-queries. Write precisely enough to be quoted whole. And a solid piece of content with real steps and real parameters does both at once.

My own routine: for a given topic, first spread the sub-query surface into H2s and subheadings (breadth), then make sure each H2 has a 130–170 word passage underneath it that can be quoted standalone (depth). Once both are done, I stop worrying about how many times AI cited me this month — the work is finished, and the rest is up to how quickly the models iterate.