Google AI Mode and query fan-out: how it splits your search

What query fan-out is, and why Google calls it "question splitting"

Query fan-out is the mechanism that lets AI Overviews and AI Mode run several related searches across subtopics and different data sources to assemble a single answer. This is not an SEO blogger's theory — it is in Google's own public documentation on "AI features and your website", whose Spanish (es-419) edition has its own name for it: división de preguntas, literally "question splitting". Search Console Help describes AI Mode in similar terms: it groups your question into subtopics and searches them simultaneously, and every follow-up question you ask within that conversation counts as a brand-new query, not a continuation of the first one.

Worth pausing here, because what is query fan out actually hides two separate questions, and conflating them is where most explanations go wrong. One is technical: how Google builds the answer internally, which is what we just described. The other is practical: what you can actually do with that knowledge, and that is where most guides stop short. Knowing fan-out exists tells you nothing actionable if you do not know which part of that machinery you can actually observe.

That is the question this post answers. We are not going to reverse-engineer Google's algorithm — nobody outside Mountain View has that data — we are going to approximate the likely cluster of subqueries with sources that genuinely exist, and then measure those subqueries like ordinary keywords: organic position, plus the AI Overview citation when one shows up. The rest of the post is exactly that path, step by step.

AI Mode vs. AI Overviews: what changes for measurement

The difference between AI Mode and AI Overviews is not just cosmetic — it is about what you can actually measure from each one. An AI Overview is a block inside the classic results page: it shows up when Google thinks it helps — per Google's own documentation, often it does not — and occupies a single position shared by every link it cites. Underneath, it is still a normal Google search with an extra block on top.

AI Mode is a different thing entirely: a separate, conversational experience with more links than an AI Overview, where every follow-up question you type counts as a new query within that session. Google says the two run on different models and techniques, so the answer and the links an AI Overview and AI Mode show can differ even for the exact same starting search, even though both can lean on fan-out to build their response.

AI Mode is already live in Spanish on Google.com across much of LATAM, but the rollout happened in phases by country and device. Before you weight it heavily in your strategy, check your own Google.com to see whether the tab or button next to the search box has actually shown up for you — that is more reliable than any generic date.

For this post, the difference that matters is this: an AI Overview can be observed keyword by keyword, because it occupies a fixed spot in a SERP you are already tracking the usual way. AI Mode has no such anchor — it lives in its own tab, its own conversation — so Search Console only gives you aggregate impressions per page, with no query and no position. That asymmetry is what drives the whole matrix in the next section.

What you can measure on each surface: the matrix

With the context from the last two sections, here is the table that sums it up: what data actually exists on each surface, where it comes from, whether it is per query, whether it carries a position, and whether Rankiamos measures it today.

The only row with an exact position, day by day, city by city, is the organic row: that is real positions by city, on mobile and desktop, and it is the foundation everything else in this post builds on.

The AI Mode row deserves one clarification: Search Console's generative AI performance report counts AI Mode inside the "Web" search type and shows page-level impressions, not per-query, combined with AI Overviews. We have already covered how to read that whole report and its limits elsewhere; here it is just one more row in the matrix.

Why nobody sees Google's real fan-out (and what the simulators are worth)

Google does not publish its subqueries. Not in Search Console, not in its documentation, not through any API. That gap got filled fast by products promising to show you the query fan out seo for your keyword: "query fan out tool", "analyzer", "generator". What most of them actually do is ask a language model to guess related subtopics. It is not privileged Google data — it is another model's opinion, trained on different data.

There is a quick way to check this yourself: run the same prompt three times in a row asking for the fan-out of one keyword, and compare the three lists. If the output barely changed, it would carry some value as an approximation. In practice it varies quite a bit between runs, because it is not retrieving a fixed data point — it is generating plausible text each time.

That does not mean these tools are useless. They serve the same purpose as any AI-assisted brainstorm: they surface angles you might not have thought of, and some of those angles will overlap with real subqueries. The mistake is treating the output as a fact instead of a hypothesis. The rule we apply at Rankiamos — worth repeating because it is the easiest one to drop under deadline pressure — is simple: a subquery only enters the cluster if a real source confirms it, never just because an LLM suggested it.

And it is worth being explicit about our own product here: the app does not build the cluster for you. Rankiamos does not generate fan-out, does not simulate subqueries, and will not tell you "these are the 8 questions Google is splitting your search into." What it does is measure — with real position and citation data — the subqueries you decide to add. Building the cluster, which the next section covers, stays human work built on real sources.

Two real sources for approximating subqueries: Search Console and the SERP

If simulators are not a real data source, what is? Two sources, neither perfect on its own, but together they give a reasonable approximation of the likely subquery cluster.

As a starting point, before you even have accumulated Search Console queries, Rankiamos's site scan — available from the Free plan — returns up to 50 keywords your domain already ranks for in the chosen country. It is not a source of fan-out subqueries, but it is a solid way to decide which main query is worth building a cluster around.

Building the cluster: from one main query to 5-12 measurable subqueries

Using the two sources above, here is the full procedure for turning one main keyword into a cluster of subqueries genuinely worth tracking.

1. Step 1 · Pick the main query

   Something with commercial value, ideally one where you already see an AI Overview or that has a comparison or decision-criteria shape. The e-commerce example we will use throughout: "best running shoes for beginners".

2. Step 2 · Pull candidates from Search Console

   Filter by the page that answers it, use a range of three months or more, and keep queries with five or more words plus question-shaped ones. When you connect Search Console for free to your rank tracker, the query list and the "ranking URL" sit in the same place, which saves you bouncing between two tabs while you build the cluster.

3. Step 3 · Add the SERP from every country you sell in

   Note down "People also ask" and "Related searches", with date and country. The same main query can trigger different questions in Mexico than in Argentina.

4. Step 4 · Use an LLM only to spot gaps

   Optional, with a strict filter: no candidate enters the cluster unless one of the real sources above confirms it.

5. Step 5 · Group by sub-intent

   Definition, cost, comparison, specific need, maintenance, alternatives. Keep one representative subquery per group: 5 to 12 total, never the 40 you could theoretically list.

6. Step 6 · Assign the URL that should answer each one

   After the first measurement, compare that assignment against the actual "ranking URL". If they do not match, that is your first finding.

7. Step 7 · Load the cluster as keywords

   Each subquery per city or country is 1 quota unit, and each lexical variant per country counts separately; on Pro, each device you measure burns its own unit.

8. Step 8 · Measure for 4 weeks before touching content

   Daily organic position and AI Overview citation if you are on Pro, plus page-level impressions from Search Console's generative AI report as a background signal.

What a finished cluster looks like, using "running shoes" (Mexico, Colombia) alongside a lexical variant, "zapatillas para correr" (Argentina, Chile, Peru):

Illustrative example for format purposes only: fill each row with the real data from your own Search Console and SERP.

How much quota a cluster costs, and how to prioritize it inside your keywords

The math on the cluster above: 7 subqueries × 3 countries (Mexico and Colombia share "running shoes"; Argentina, Chile and Peru use the "zapatillas" variant, but every lexical variant still counts on its own) comes to 21 quota units. With Pro's 50 units, that fits roughly 2 full clusters of that size with daily checks. With Free's 6 keywords and a weekly check, a cluster that size does not fit whole: the free plan is enough for a smaller, one-country pilot of a cluster — fine for a general trend read, not for watching the whole market.

With limited quota, not every subquery in the cluster deserves the same spot in line. Four criteria, in this order:

One warning on criterion 2: search volume measures what people type into a keyword tool, not what AI Mode queries internally when it splits your question. Use it to rank candidates against each other, never to drop a subquery your other sources already confirmed just because its reported volume comes back as zero or "no data".

What to measure per subquery: organic position and the AI Overview citation

With the cluster prioritized and loaded as keywords, three things are worth checking week over week per subquery: daily organic position, whether an AI Overview appears and cites you — the citation piece being AI Overview citations included in Pro — and the "ranking URL", to confirm whether the page you assigned in step 6 is actually the one answering.

As a background signal, with no per-subquery breakdown, there are the page-level impressions from your Search Console's generative AI report. They do not replace per-keyword measurement, but they confirm whether that page's overall exposure across AI surfaces is trending up or down over time.

With that data you can calculate a rough cluster coverage: what share of subqueries sit in the top 10 and what share carry a citation in their AI Overview. It is a useful number for comparing one cluster against another, or one month against the next, but do not confuse it with "presence in AI Mode" — it is still a measurement of the observable surfaces, not of AI Mode itself.

If within the cluster you spot a subquery where you sit in the top 10 but the AI Overview does not cite you, that is exactly the scenario we cover in more depth — the tracking method and the analysis of who it cites instead — in the guide on how AI Overview citations get tracked.

With four weeks of cluster data, four readings come up often, along with what to do about each:

One point worth more than any trending tactic: Google is explicit that AI Overviews and AI Mode do not require special optimization beyond the usual SEO fundamentals — clear, well-structured content that answers real intent. Google AI Mode SEO and AI mode SEO are not a new discipline with its own rulebook; it is the same SEO as always, applied to subqueries you now know how to approximate.

And it is worth closing where we started: this does not measure your presence in AI Mode; it measures the measurable surfaces for the likely subqueries. Everything you did in this post — building the cluster, loading it, tracking position and citations — is exactly that: the best approximation available today, with Search Console, site scanning and daily tracking in one app, not a direct measurement of something Google still does not expose.