Query fan-out SEO, turning a buyer prompt into a content plan

Query fan-out is a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results before answering. That is the definition Google publishes in its generative AI optimization guide, last updated in July 2026. The SEO consequence is immediate, one user question triggers several searches, so several different pages can feed the same answer. Google states elsewhere that the technique applies to AI Overviews as well as AI Mode, and that it supports greater link diversity than a traditional search. In practice, a page that only answers the head phrasing forfeits the subqueries deciding the rest of the response, while a set of pages covering each sub-intent multiplies the possible entry points into your site.

Google pairs that definition with a concrete example. If the original query is how to fix a lawn that is full of weeds, fanout queries might include "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn". Three sub-intents, three editorial angles, three pages that can each be cited, where classic SEO would have targeted a single head keyword.

One clarification matters immediately, because it changes the scope of this guide. The term query fan-out belongs to Google, the mechanism does not. OpenAI, Anthropic and Perplexity each document, in their own words, a model that writes its own queries to answer. We cover those documents further down with the exact wording. For now, note that a content plan built on this principle does not serve Google alone.

Query fan-out does not replace search optimization, it changes the unit of work. You no longer target one query and one page, you target a user question and the whole set of follow-up questions the model asks to answer it. Google makes the same point in its AEO and GEO callout, optimizing for generative AI search is optimizing the search experience, and therefore still SEO.

Google, generative AI optimization guide | Google, AI features and your website

Related concepts. Prompt intent | AI Overviews

Start from a real buyer question rather than a keyword. We submitted this prompt to AI Mode, which helpdesk software should a 20-person B2B SaaS choose with EU hosting and GDPR compliance. That phrasing is representative of what buyers actually type today, a team size constraint, a geographic constraint, a regulatory constraint and a purchase decision, all in one sentence.

The answer is not a list of links. It is a structured document that answers several questions nobody asked out loud. It names three options in the opening lines, it separates European vendors from US vendors offering an EU region, it builds a comparison layer, then it closes on contractual checks to run before signing. Each of those blocks cites different sources.

A buyer prompt always breaks down along the same seven axes. This grid does not claim to reproduce the internal queries Google fires, it reproduces the structure of the decision, which is enough to build a useful content plan. We apply it to the European helpdesk example, but it works for any B2B category.

Three official publications establish the facts about query fan-out. In March 2025, launching AI Mode, Google described a technique that issues multiple related searches concurrently across subtopics and multiple data sources, then brings those results together into one response. In May 2025 at Google I/O, Google added that AI Mode breaks a question into subtopics and issues a multitude of queries simultaneously on the user behalf, and that Deep Search pushes the same technique to hundreds of searches to produce a fully cited report. The Search Central documentation then gives the formal definition and lists it as one of the two mechanisms behind generative AI search, alongside retrieval-augmented generation, also known as grounding. Nothing in those three sources describes how many subqueries a given prompt produces, or which ones.

The Search Central documentation industrializes that message. The AI features page states that both AI Overviews and AI Mode may use the technique, and that it supports greater link diversity than a traditional search. The generative AI optimization guide adds the formal definition and the weeds example quoted above. Both pages also set the technical bar, your page must be indexed and eligible for search snippets, and the nosnippet, data-nosnippet, max-snippet and noindex directives remain the available controls.

Read the verb Google uses carefully. The documentation says these features may use query fan-out, not that they run it on every query. Presenting fan-out as a universal, deterministic mechanism would be an extrapolation. What Google does confirm is that the technique exists, what it is for, and that it widens the number of pages that can be cited inside one answer.

Google, AI Mode launch, March 2025 | Google, AI Mode and Deep Search, May 2025 | Google, AI features and your website

No. The name is specific to Google, the behaviour is shared by every answer engine that documents its search layer. OpenAI writes that ChatGPT search typically rewrites your query into one or more targeted queries sent to its search partners, and that it may send additional, more specific queries after reviewing the initial results. Its example is as telling as the Google lawn one, a question about drugs targeting CCR8 in oncology first becomes "CCR8 immunotherapy drug development 2025", then "CHS-114 conference 2025" once the first results are read.

Anthropic describes the same loop on the API side. The web search tool documentation states that Claude decides when to search, that the API runs the searches and returns the results, and that this process can repeat multiple times throughout a single request. It even gives an order of magnitude rarely published elsewhere, simple factual queries typically use 1 to 3 searches, while comparative or multi-entity research can use 10 or more. That is query fan-out, quantified by the provider itself.

Perplexity built it into its API. The Search API accepts up to five related queries in a single request, precisely to explore different angles of a topic, with each query processed independently. The vocabulary changes from one provider to the next, the editorial consequence does not, a buyer question is answered with a set of pages rather than a single page.

OpenAI, searching the web with ChatGPT | OpenAI, web search tool | Anthropic, web search tool | Perplexity, Search API | Google, grounding with Google Search

Guides by engine. Get cited by Perplexity | Get cited by ChatGPT

Start with the one genuine blind spot, AI Mode and AI Overviews. Google does not expose the subqueries behind a question there, neither in the displayed answer nor in Search Console. The report dedicated to generative AI features, announced in June 2026, measures your impressions and breaks them down by page, country, date and device. It carries no query dimension, and it is rolling out to a subset of site owners. In other words, the closest official tool tells you that you were shown, never which subquery caused it. Anyone selling you the exact list of AI Mode internal queries is therefore selling an estimate rather than a measurement.

The rest of the landscape is far more open, and that is the good news in this guide. ChatGPT, Claude and Perplexity display the searches they run, right in their interface, while they answer. You only have to read them. On the API side, an Anthropic response contains a server_tool_use block holding the query the model wrote, the OpenAI web_search_call output usually includes the queries actually run, and the Gemini documentation states that the google_search_call block contains the search queries the model executed. For Gemini that API route is the only one, the consumer interface does not show them.

The dividing line is therefore not Google against the rest, it runs between surfaces. On AI Mode and AI Overviews you stay blind. On the other engines you can read the subqueries one by one. The real work is no longer obtaining them, it is collecting them regularly, against a stable prompt set, so a one-off rephrasing can be told apart from a durable pattern.

One distinction stays useful when comparing tools. Logging a query displayed by ChatGPT, Claude or Perplexity is a measurement, the query comes from the engine. Inferring what AI Mode might have searched is an estimate. Both have their place, they do not deserve the same wording, and mixing them in one table produces fragile conclusions.

We ran that test on our own example. Asking a narrow follow-up about CLOUD Act exposure at the same time, AI Mode produced a decisive answer built on an entirely different source set. Not one domain cited on the buying prompt reappeared on the subquery. On a verifiable case, that confirms a sub-intent is won with its own page, not with a paragraph buried inside a general article.

• Never present a third-party subquery list as the real set of queries Google fired.
• Date every observation, an AI Mode answer is not reproducible verbatim.
• Test English and French separately, our two captures return different brands and different sources for the same intent.
• Never conclude from a single test, repeat the measurement before making an editorial decision.

Google, generative AI performance reports in Search Console | Search Console, generative AI performance report | Anthropic, web search tool | OpenAI, web search tool

Related Hikoo guides. Which prompts mention your brand | Find the sources shaping LLM answers

A content plan built for query fan-out looks like a hub surrounded by satellites. The hub answers the full buying question, carries the comparison table and commits to a recommendation. Each satellite covers one sub-intent in depth, with a self-contained answer in its opening lines. The hub links to every satellite, each satellite links back, and satellites that share a criterion cite each other.

This table maps every sub-intent found in the AI Mode answer to the editorial format that serves it, the proof it requires, the target page and the internal link that ties it back to the cluster. Fill it once per buyer prompt and it becomes the production brief.

Related concepts. Internal linking | Semantic search

A fan-out grid always produces more topics than you can write. Useful prioritization cannot run on estimated volume, since nobody has volume data for internal subqueries. It runs on three signals you can actually observe, how often the sub-intent shows up in the answers you tested, how strong the currently cited sources are on that sub-intent, and how close the sub-intent sits to what you sell.

In practice, a blocking sub-intent that appears across several answers, is currently served by thin or ageing pages, and sits close to your offer, goes first. Conversely, a sub-intent already covered by solid institutional sources needs a genuinely new angle, otherwise it burns time without moving your position. Google says as much in its guide, commodity content adds little unique insight, and it is the distinctive viewpoint that pays off over time.

• Score recurrence, how many of your tests surface this sub-intent inside the answer.
• Score the weakness of incumbent sources, freshness, depth and authority of what is cited today.
• Score commercial proximity, does the sub-intent genuinely move the reader towards a decision.
• Handle blocking objections first, they are underserved and they unlock stalled buying cycles.
• Re-score the grid every quarter, answers shift and cited sources change.

Go further. Competitive analysis of AI citations | Elevate

Measuring query fan-out work runs on three complementary planes, none of which is enough alone. The first is Search Console, with the generative AI performance report for how your impressions inside those features move, and the standard report for web search. The second is prompt tracking, replaying the same buyer questions on a fixed cadence and logging the brands and sources cited. The third is source analysis, knowing which content, yours and everyone else, actually feeds the answers in your category. None of these three planes returns the subqueries themselves, they return effects, which is enough to arbitrate an editorial plan but never enough to reconstruct the engine reasoning. Run all three or the picture stays partial, since each one is blind to what the other two see.

One methodological point avoids a lot of disappointment. Gains read at cluster level, not at page level. A satellite may never be cited directly while still making the hub citable, because it supplies the depth and the entities that were missing. Conversely, a rise in generative AI impressions says nothing about the subquery behind it, since that dimension does not exist in the report.

• Freeze a stable list of buyer prompts and replay it on a fixed cadence, otherwise your comparisons mean nothing.
• Track the share of answers where your brand appears, rather than a rank, which has no meaning inside a generated answer.
• Log the queries the engines display too, they hand you the entities to cover instead of leaving you to guess.
• Cross cited sources against your own pages to spot the sub-intents where you are absent.
• Log every satellite publication with its date, so a citation change can be tied back to an action.

Hikoo resources. Google Search Console for AI | AI Overviews impressions in Search Console | Appear in Google AI Overviews | Spotlight

Most mistakes come from a magical reading of fan-out. Google published a full myth-busting section in its optimization guide, and it contradicts several widespread practices head on. llms.txt files are not used by Google Search. Breaking content into tiny fragments is not a requirement. Structured data is not required for generative AI search and no special schema.org markup exists for it. Rewriting a text to match an exact phrasing is pointless, since the systems understand synonyms.

A second mistake is promising mechanical results. Semrush published an experiment where four articles were expanded to answer fan-out queries. Citations went from two to five in one month, but over the same period the measured share of voice fell from 23.4 percent to 20.0 percent and brand visibility from 13.6 percent to 10.6 percent. Even advocates of the method publish mixed results, which is a good reason to treat fan-out as a coverage discipline rather than a lever with guaranteed effect.

• Believing that covering subqueries guarantees a citation, no official source claims it.
• Publishing one thin page per subquery, which produces exactly the commodity content Google says it values least.
• Confusing subqueries inferred by a third-party tool with the real queries Google issued.
• Ignoring the update date, when freshness is often the only thing separating two equivalent pages.

Google, generative AI search mythbusting | Semrush, query fan-out experiment, September 2025

Pick one buyer prompt, the one your sales team hears most often. Submit it to AI Mode, log the sub-intents and cited sources, fill the grid from this guide, then produce the hub and two satellites. Within a quarter you will know whether your category rewards depth or freshness, and you will hold a repeatable method rather than an intuition.

To follow the effect of that work without guessing, Hikoo replays your prompts, logs the search queries the engines actually run alongside the brands and sources cited, and shows where your cluster still misses a sub-intent. The 126 variants quoted above come from exactly that. Same material as the one used in this guide, measured continuously rather than once.