One Query Becomes 12 Searches — Sub-Query Coverage for AI Search Visibility (2026)

Key takeaway: When a user types one question into AI search, the engine does not search that question as-is. Dounia Berrada, Google's Senior Engineering Director for Search, confirmed it on the official blog: "AI Mode performs 12 searches per query by default" (Google, March 2026). The unit of visibility competition has shifted from "the question the user typed" to "twelve invisible sub-queries the engine generates." This article explains what that structure changes for brands, together with the two-level branching pattern sub-queries follow (as of 2026-08-06).

---

Three-line summary

• The unit of competition has changed. AI Mode splits one question into 12 sub-searches by default (officially confirmed by Google). Even if you rank first for the user's question, you miss the answer-material shortlist if none of your pages responds to the twelve sub-queries.
• Sub-queries branch in predictable patterns. A first level of intent axes — types, comparisons, pricing, reviews — and a second level of attribute axes — features, segments, region, definitions. The actual lists are private, but the branching patterns can be mapped.
• The response unit is coverage, not keywords. "One page, one keyword" thinking targets a single cell out of twelve. How much of the sub-query space your content can answer determines your odds of appearing in the synthesized answer.

The officially confirmed facts

Query fan-out is not speculation — Google has documented the behavior repeatedly. In chronological order:

When	Officially confirmed	Source
May 2025 (I/O 2025)	AI Mode breaks a question into subtopics and issues many searches at once; Deep Search scales to hundreds	Google blog
Sept 2025	Visual search fan-out — recognizes main subjects plus subtle details and secondary objects in an image, running multiple queries in the background (English, U.S. rollout)	Google blog
Mar 2026	"12 searches per query by default" — stated by the Senior Engineering Director for Search	Google blog
May 2026 (I/O 2026)	AI Mode passed 1 billion monthly users in 12 months	Keynote coverage

"AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf." — Google, I/O 2025 official announcement

ChatGPT belongs to the same family. Industry reporting puts its branching at 4–20 sub-queries depending on complexity (Profound 2025-10, via Ekamoira) — an external estimate, so treat it as reference rather than fact.

What it means — the visibility contest happens out of sight

The practical meaning of these numbers comes down to one thing: your content competes not on the question the user typed, but on the sub-queries the engine generated internally.

In traditional search, optimizing one page for "accounting service for small businesses" settled the contest on that keyword's SERP. Under fan-out, the same question internally splits into sub-searches like "types of accounting services," "tax accountant vs accounting SaaS," "accounting service pricing," and "accounting reviews for solo founders" — and the AI pulls the passages it needs from each sub-search (retrieval) and composes them into a single answer. A page that never ranks for the user's question can still be selected as answer material through some sub-query; a page that dominates the user's question can still be dropped if it covers none of the branches.

Visual search fan-out signals that this structure is expanding beyond text. When secondary objects inside an image each become queries of their own, the contextual metadata of product images and visual assets — alt text, captions, structured data — becomes answerable material for sub-queries too.

How sub-queries branch — first and second levels

The actual sub-query lists are private, but public examples and measured reports converge on a two-level map. Here is an example with "recommend an accounting service for small businesses" as the representative question.

Level	Axis	Example sub-query
1st (intent)	Types	"What types of accounting services exist"
1st	Comparison	"Tax accountant bookkeeping vs accounting SaaS"
1st	Pricing	"Accounting service price range"
1st	Reviews	"Small business accounting service reviews"
2nd (attribute)	Segment	"Accounting service for solo founders"
2nd	Feature	"Services with automated tax invoicing"
2nd	Region/market	"Local tax law support"
2nd	Definition (informational)	"What is bookkeeping outsourcing"

The first level closely mirrors the intent types users actually express across a search journey — category exploration, comparison, near-purchase. In other words, a brand that has covered intent-diverse entry questions broadly already covers much of fan-out's first level. The gap usually opens at the second, attribute level — and note that informational sub-queries like "what is bookkeeping outsourcing" are not a place where your brand gets named, but a place where your page gets selected as answer material. In those cells, content existence — not brand mentions — decides the outcome.

The shift for brands — from keywords to coverage

The practical implication reduces to one change of unit. Where traditional SEO optimized "keyword × page," the fan-out environment rewards "sub-query space × coverage." Defending the #1 spot on a single keyword defends one cell out of twelve; if the other eleven are empty, you can lose the synthesis.

The content-format implication is also singular — for one page to be selected across multiple sub-queries, it must be divided into self-contained passages (a direct answer right after a question-style H2) that still make sense when extracted alone. Pages that extract well beat pages that merely read well.

Which sub-queries are empty, and where you are already being selected, is a matter of measurement rather than intuition — AI answers shift over time, so repeated measurement is the premise. Measurement tools such as RanketAI offer visibility opportunity analysis, which auto-discovers customer entry questions by intent type and checks whether your brand appears, and AI brand visibility analysis, which repeatedly measures citations and mentions on your key questions.

Frequently asked questions

Can I see the actual sub-queries an engine generates?

No. Neither Google nor OpenAI exposes the sub-query lists. Even "12" is a statement about the default count, not a list. Every coverage design is therefore an approximation based on branching patterns — and the practical value lies in reducing empty cells, not in the precision of the approximation.

Should I abandon "one page, one keyword"?

It needs to evolve. Under fan-out, a page that can answer several sub-queries has better odds of being selected. But the opposite extreme — cramming every intent into one page superficially — is also risky, because the page then matches no sub-query strongly. A realistic balance is two to four core sub-questions addressed head-on per page.

How is this different from long-tail keyword SEO?

The target space overlaps, but the trigger differs. Long-tail SEO only works when a user actually types that query. Sub-query coverage also works when the user types only the representative question, because the engine runs the sub-queries on their behalf. The same content earns selection opportunities from both direct searches and internal branching — that is the difference.

Where should I start with no budget?

Restructuring existing pages gives the best return. Before producing new content, reorganize your key pages into question-style H2s with self-contained answer passages — the same page can then respond to more sub-queries. If new production is needed, informational content covering definitions and how-tos is usually the cheapest to produce and the fastest to enter answers.

Do I need to respond to visual search fan-out now?

Non-English markets can afford to watch — the rollout is English and U.S. only so far. But the low-cost preparation — alt text, captions, and structured data for product and service images — overlaps with existing accessibility and SEO work, so it is worth doing in advance.

Related reading

• Google AI Mode Optimization — llms.txt and 7 Signals — the technical signals that get you into AI Mode's answer-material shortlist.
• LLMs Build Answers in 4 Stages — where fan-out sits in the full answer-generation pipeline.
• Content Design for the Conversational Query Era — how to cover multiple sub-queries in one page with question-style H2s and FAQs.