Almost everything written about query fan-out is aimed at enterprise SEO teams with keyword lists and simulation tools. Almost none of it is aimed at the people it will actually affect first: the restaurant, the shop, the plumber. This piece is the second version. Same mechanism — Google's AI Mode splits every question into 8 to 12 parallel searches — explained from behind the counter, with no tooling required.
Why this is worth ten minutes of your time now
On 22 July 2026, Google switched on AI Overviews and AI Mode in France — the last major market to flip, after two years held up by a dispute with French press publishers over neighbouring rights. Every other significant market went first.
Which makes the timing useful rather than academic. If you operate in a market that flipped a year ago, the behaviour described below is already shaping your traffic. If you're in one that just flipped, you get to watch it happen with a baseline still intact. Either way the mechanism is the same, and hardly anyone running a small business has looked at it.
What AI Mode actually is
AI Mode is a full search tab, sitting alongside Images, News and Maps. You ask a question in plain language, you get a written answer with a few sources, and — the important part — you can follow up: refine, correct, dig deeper, across one continuous exchange. It's also reachable directly at google.com/aimode, and it requires being signed in to a Google account.
The distinction from AI Overviews is worth holding onto. An AI Overview is imposed: it appears above the normal results whether you asked for it or not. AI Mode is chosen: you decide to go in. The first will move everyone's numbers quickly. The second builds a habit more slowly — but it's the clearer signal of where Google intends search to go.
The mechanism: query fan-out
When you ask AI Mode a question, Google doesn't go looking for "the" page that answers it. It applies a technique called query fan-out: it parses the question, identifies the entities and constraints inside it, then breaks it into several sub-queries which it runs simultaneously against its index. It pulls answer fragments for each, then writes one synthesis, citing the sources it kept. Google says this runs on a custom version of Gemini built specifically for that decomposition.
The scale: analyses published through 2026 commonly observe 8 to 12 sub-queries for an ordinary question. For a complex one, the Deep Search feature can fire off several hundred.
A worked example
Take a question a customer might genuinely ask: "where can I get lunch near Bastille with a terrace, vegetarian, not too expensive, open on Monday?"
Classic search would have looked for that phrase. AI Mode instead fires, in parallel, something like:
- restaurants near Bastille with a terrace
- vegetarian restaurants Paris 11
- lunch price range Bastille area
- restaurants open Monday Paris 11
- recent reviews vegetarian Bastille
- vegetarian lunch menus Paris 11
Then it cross-references and recommends two or three places. Look closely at what follows from that: you don't need to be "the best result" for the whole question. You need to answer several of those sub-questions cleanly. A restaurant whose website mentions its terrace nowhere, says nothing about vegetarian options and doesn't list opening days is eliminated on three of six sub-queries — whatever its Google ranking says.
Two assumptions this breaks
You no longer optimise a page for a keyword. You prepare it for a cluster of sub-questions you can't see and that Google generates on the fly. Picking a primary keyword and repeating it loses most of its meaning.
The unit selected is no longer the page — it's the passage. AI Mode doesn't rank ten pages; it picks fragments, each answering one sub-query. A long, excellent page whose useful fact is buried inside a ten-line paragraph loses to an average page that answers in two clean sentences under an explicit heading.
Why this quietly favours small businesses
This is the part the enterprise-facing coverage tends to miss, and it's the reason I think fan-out is good news if you run a single location.
Fan-out sub-queries are narrow and factual. Does it have a terrace. Is it open Monday. Do they do gluten-free. What's the price range. Is it accessible with a pushchair. How long do alterations take. These aren't questions you win with domain authority or a content budget — they're questions you win by knowing the answer and writing it down.
A national directory answers those vaguely, because it aggregates thousands of businesses and can't be specific about any of them. You can be exact. On a sub-query that precise, exactness beats authority. That is a genuinely unusual opening, and it will not stay open indefinitely.
The catch is the obvious one: the information has to exist, in writing, on a page a crawler can read. Knowing it doesn't count. Saying it on the phone doesn't count.
Six levers
1. Map the sub-questions from your phone calls
Skip the tooling. For two weeks, write down every question customers ask you on the phone or at the counter. That list is your fan-out map, and it's more accurate than anything a keyword tool would generate, because it comes from the actual demand rather than an inferred one. Every recurring question deserves an explicit answer on the site.
2. Write self-contained passages
Every paragraph must stay true and intelligible once pulled out of context. No "as mentioned above", no "in that case" without restating which case. Repeat the subject instead of using a pronoun. Slightly less elegant read end to end; considerably more effective at extraction.
3. Structure on the lines where the model cuts
The retrieval layer segments on headings, lists and table cells. Headings phrased as real customer questions, short paragraphs, lists, tables — these aren't web-writing mannerisms, they're the cut lines. A well-segmented page produces cleaner candidate passages.
4. Supply what an AI can't generate
Models discard generic content they could have written themselves. What holds them: dated and sourced figures, first-hand data, concrete cases, demonstrable experience. One real price, one real lead time, one constraint only a practitioner knows, beats ten paragraphs of general advice.
5. Consolidate your entity
AI Mode reasons over entities, not strings. It needs to know who you are, consistently from one source to the next: same name, same address, same description on your site, your business profile, directories and local press. Structured data declares that identity unambiguously; NAP inconsistencies destroy it.
6. Guarantee technical access
Static or server-side rendering — JavaScript-injected content is read poorly, when it's read. AI crawlers allowed in robots.txt. Decent speed. Nothing new, but it's the precondition: a page the system can't read is a candidate for no sub-query at all.
Test it on your own business
Ten minutes, and more instructive than most audits. Open AI Mode and ask the three or four questions a customer would ask before choosing you — phrased the way they'd phrase them, not the way an SEO would write them. See who gets cited. Follow up, the way a real user would.
Note two things: which businesses come out, and above all what's missing from the answer about you. Nine times out of ten the absent information exists — it's just buried in a page, or nowhere on the site. That's your task list.
Take screenshots. Generative answers drift over time; with no record you'll have nothing to compare against in three months. For the numbers, Search Console now exposes your impressions inside AI Mode — I've covered how to read that report and what it leaves out.
Frequently asked questions
What is query fan-out?
The technique by which AI Mode breaks a single question into several sub-queries, runs them in parallel against Google's index, then synthesises one sourced answer. Analyses in 2026 commonly observe 8 to 12 sub-queries; Deep Search can fire hundreds.
Does it matter for a small local business?
Yes, arguably more than for a large brand. Sub-queries are narrow and factual, and on questions that specific a small business that answers plainly beats a directory that answers vaguely — provided the information is written on the site.
Do I need a keyword tool?
No. The questions your customers ask you on the phone are a better and freer source of sub-questions than any simulation tool.
Should I create one page per sub-question?
No. Thin pages are what models discard. One solid page covering a topic completely, cut into clean sections that each answer a sub-question, performs better.
Is classic SEO still useful?
Yes, as the foundation. AI Mode draws on Google's index — a page absent from it is a candidate for nothing. What changes is the unit of measurement, from page ranking to passage selection.
Want to know what AI Mode says today when a customer searches for what you do? Ask for a diagnostic: I run your customers' real questions, show you who gets cited instead of you, and what's missing on your site to be there. That's the core of my AI search optimization work.