Query Fan-Out vs. the Keyword Era: 9 Things I Learned About the Hidden Searches Behind Every AI Answer

The one search that turned into 420

Here is the number that reset my thinking. When Despina Gavoyannis at Ahrefs asked ChatGPT Deep Research to help buy a red phone case, the model ran 420 searches and cited 30 sources. Google's AI Mode, faced with the same shopping intent, typically fires 5 to 11.

Neither number is one. That is the entire story.

Google confirms the mechanism in its own documentation. Both AI Overviews and AI Mode may use a query fan-out technique, described as <a href="https://developers.google.com/search/docs/appearance/ai-features">"issuing multiple related searches across subtopics and data sources"</a> while the response is being generated. The stated payoff for the user is breadth: Google says the technique lets it surface a wider and more diverse set of supporting links than classic web search would.

For us as content creators, that breadth is both the threat and the opening. Below are the nine things we learned, in the order they changed how we work.

 

---

 

1. The retrieval model flipped from many-to-one to one-to-many

Classical information retrieval was one-to-one. You typed a string, an inverted index matched that string, and you got a ranked list.

Then semantic search made it many-to-one. Google understood that "Sydney plumber" and "plumbing service in Sydney" resolve to the same intent, so many phrasings collapsed into one result set.

AI search inverted the arrow. It is now one-to-many: one prompt is decomposed into many synthetic sub-queries, each retrieved separately, then fused back together.

Gavoyannis puts the consequence bluntly: <a href="https://ahrefs.com/blog/query-fan-out/">"Ranking #1 for a single query isn't enough anymore."</a> I resisted that sentence for about a week. Then I audited our own citations and found we were being pulled into AI answers for phrasings we had never targeted, and skipped for the exact phrase we owned.

2. There are seven recognizable fan-out formats, and you can predict most of them

Fan-out looks random until you group it. Drawing on Google's patent filings, Mike King at iPullRank catalogued the synthetic query types that recur across AI Mode, ChatGPT, and Perplexity. Ahrefs mapped them into seven working buckets, and in our testing they hold up well.

1. Related topics. "Meal prep for beginners" spawns searches for containers, easy recipes, and storage.

2. Implicit questions. "Switching to solar panels" spawns cost, installation time, and payback period.

3. Comparative queries. "Project management software" spawns Asana versus Monday, tools for small teams, pricing comparisons.

4. Recency. "Best smartphones" spawns the current year, latest releases, top rated this month.

5. Reformulations. "How to reduce bounce rate" spawns improve engagement, keep visitors longer, decrease exit rate.

6. Contextual variations. "Best restaurants" spawns your city, your usual cuisine, open now.

7. Next-step queries. "Symptoms of diabetes" spawns diagnosis, treatment options, diet plans.

A concrete example makes it click. Ask ChatGPT how to start an SEO podcast and it does not search that phrase. Ahrefs logged the actual background queries, which included podcast naming and branding ideas, technical setup, best podcast hosting and distribution, guest sourcing in marketing and tech, promotion through SEO and social, and even "podcast recording equipment."

Read that list again as a content brief. If your podcast guide covers structure and nothing else, you are eligible for roughly one query out of eleven.

3. Under the hood it is a six-step pipeline, and step four is where pages get killed

The technical process is less mysterious than the marketing chatter suggests. Ahrefs breaks it into six stages, and the vocabulary is worth learning because it tells you exactly where you can intervene.

1. Query analysis. The model classifies intent, complexity, and required response type in milliseconds.

2. Decomposition. The prompt is split into sub-queries covering each unresolved dimension.

3. Parallel retrieval. Sub-queries hit web indexes, knowledge graphs, product databases, and specialized repositories simultaneously.

4. Synthesis via reciprocal rank fusion. Multiple ranked lists are merged into one using RRF, which rewards documents that appear consistently across lists.

5. Scoring. A document ranking #2 on one list and #5 on another accumulates roughly 1/2 + 1/5. Documents present in several lists compound their score.

6. Final ranking. The fused set is re-ranked, and the winning passages become the grounding context for the generated answer.

Step four is the one I want tattooed on the wall. Reciprocal rank fusion rewards consistency over peak position. A page ranked fifth for six sub-queries can outscore a page ranked first for one. Surfer SEO's study reached the same conclusion from the data side: ranking across multiple fan-out queries raises your odds of being cited.

That single mechanic explains why comprehensive articles beat surgically targeted ones in AI answers, and why our thin pages died quietly.

4. Fan-out depth is measurable, and ambiguity is the dial

We assumed fan-out volume was arbitrary. It is not. Research from Seer Interactive and Nectiv, the latter based on an analysis of 60,000 Google fan-out queries, found an average of 9 to 11 sub-queries per prompt. Around 59% of prompts trigger 5 to 11 searches, 24% trigger 12 to 19, and the tail reaches 28.

The variable driving depth is unresolved context. Underspecified prompts leave the model with two options: ask the user, or go find out. Claude asks clarifying questions upfront and consequently runs fewer background searches. ChatGPT Deep Research does the opposite and brute-forces the ambiguity, which is how "buy red phone case" became 200 searches on phone model and case type alone before it moved on to anything else.

The practical read for us: the more implicit context your topic carries, the more surface area exists for you to be found. High-ambiguity topics are not harder. They are wider.

5. Fan-out queries are not long-tail keywords wearing a costume

This was our most expensive mistake. We pulled a list of observed fan-out queries and briefed writers to target them verbatim. It did almost nothing, because these strings have five properties that ordinary keywords do not.

• Synthetic. They are generated by a model, not typed by a person.

• Inconsistent. The same prompt yields different fan-outs across models and users.

• Probabilistic. Same prompt, same model, same user, still different queries on a second run.

• Context-rich. The model appends modifiers no human would ever type.

• Zero volume. Over 95% receive no recurring searches, per Seer Interactive.

For scale on that last point: Ahrefs discovers over 110 billion keywords and filters down to the 28.7 billion worth tracking. Most fan-out queries never make the cut, because nobody searches them.

So we stopped treating the list as targets and started treating it as diagnostic output. The individual strings are noise. The patterns they reveal are the signal.

6. The pattern determines the play, and the plays are different

Once we clustered fan-outs by pattern, optimization stopped being one activity and became six.

Entity-heavy topics (products, tools, services with many attributes) need explicit attribute coverage and structured data. This is where Gavoyannis's concept of feature stacking earns its keep: the mental checklist a buyer assembles but never types. For the red phone case, ChatGPT surfaced implicit criteria including specific shades of red, anti-yellowing resistance, and wireless charging alignment. None of that appears in the query. All of it appears in the research. If your product page does not state it, you cannot be matched on it.

Journey-heavy topics (complex purchases, unfamiliar categories) need clusters spanning awareness, education, comparison, decision, and implementation. "Buy laser cutter" fans out across all five stages simultaneously.

Trust-heavy topics (YMYL, high cost, irreversible decisions) trigger searches for credibility markers: author credentials, third-party citations, reviews, awards, transparent methodology, published policies. A $15 purchase needs almost no verification. A financial advisor needs all of it.

Comparative, personalized, and recency-driven patterns each get their own treatment: side-by-side evaluations with explicit decision criteria, local and contextual angles, and temporal qualifiers with genuine freshness behind them.

The trap is treating all six the same way. We were writing journey-style blog posts for entity-style queries and wondering why nothing landed.

7. The peer-reviewed data on what earns a citation is unusually actionable

Most GEO advice is vibes. One paper is not. In GEO: Generative Engine Optimization, presented at KDD '24, Aggarwal and colleagues from Princeton and IIT Delhi built GEO-bench, a benchmark of 10,000 queries across 25 domains, and tested nine content modifications against it. Their headline finding: <a href="https://doi.org/10.1145/3637528.3671900">"including citations, quotations from relevant sources, and statistics can significantly boost source visibility."</a>

The numbers, measured on position-adjusted word count and subjective impression:

Three things jumped out at us.

First, keyword stuffing actively backfires. The classical SEO reflex is now a liability, which makes sense once you accept that retrieval is semantic rather than lexical.

Second, fan-out plus GEO disproportionately helps underdogs. In the paper's multi-site experiment, adding citations lifted the fifth-ranked source by 115.1% while the top-ranked source lost 30.3% of its visibility. Backlink authority, the moat that protected incumbents for twenty years, matters less once a generative model is reading the actual text.

Third, combinations compound. Pairing fluency optimization with statistics addition beat any single strategy by more than 5.5%.

8. Optimization now happens off your site as much as on it

Many fan-out queries are searches for external validation: review sites, "best of" listicles, industry publications, comparison sites, community threads. You cannot rewrite those pages. You can decide whether you appear in them.

We now run a two-column audit. Column one: where are we already cited by third parties. Column two: where do competitors appear and we do not. The second column became our outreach list, prioritized by which fan-out patterns those sources actually satisfy. Trust signals get pulled from across the web, not just from your domain, so an unclaimed profile on a comparison site is a real gap and not a nice-to-have.

9. The metrics changed, but the fundamentals did not

Google is direct about this, and I think the industry has partly ignored it: there are no additional technical requirements to appear in AI Overviews or AI Mode. No AI text file, no special schema, no new markup. A page needs to be indexed and eligible to show with a snippet. That is the bar.

What Google does recommend is unglamorous and familiar: allow crawling in robots.txt and at the CDN layer, keep important content in textual form, use internal links so content is findable, make structured data match the visible text, and keep Merchant Center and Business Profile data current. Google also notes that clicks from result pages containing AI Overviews tend to be higher quality, with users spending more time on site.

What we changed was measurement, not fundamentals:

• Track citation frequency across platforms, not just rank. Fan-out means you get cited for queries you never targeted.

• Group pages into topic portfolios and watch aggregate movement instead of individual positions.

• Tag priority fan-out queries that do have search potential in rank tracking, and ignore the 95% that do not.

• Watch Search Console's Performance report, where AI Overviews and AI Mode impressions are folded into the Web search type.

Rankings, traffic, and conversions still matter. Citation and coverage metrics sit on top of them as a second layer.

 

---

 

What I would do first, tomorrow morning

If I were starting from zero, I would pick one commercially important topic and run four steps in order.

1. Observe the fan-out. Run the topic through AI Mode, ChatGPT, and Perplexity and log the background queries. Look for the pattern, not the strings.

2. Classify the pattern. Entity-heavy, journey-heavy, trust-heavy, comparative, personalized, or recency-driven. The classification dictates everything downstream.

3. Audit against it. Which angles do you cover, which are missing, and which live on third-party sites you do not control.

4. Apply the GEO levers that have evidence behind them. Add credible citations, add real statistics, add attributed quotations, improve fluency, and delete anything that reads like keyword stuffing.

The honest summary is that query fan-out did not invent a new discipline. It exposed something that was always true and merely unmeasurable: searchers carry a stack of requirements they never type. Traditional search made them spell it out. AI search infers it and goes looking on their behalf.

Our job is no longer to match the words in the box. It is to be the source that satisfies the twelve questions behind it.

 

---

 

Sources and further reading

• Google Search Central, AI features and your website

• Despina Gavoyannis, What is Query Fan-Out?, Ahrefs, March 2026

• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande, GEO: Generative Engine Optimization, KDD '24

• Mike King, How AI Mode Works, iPullRank

• Seer Interactive, Gemini 3 query fan-out research

• Nectiv, We analyzed 60k Google fan-out queries

• Surfer SEO, Query fan-out impact study

• Microsoft, Reciprocal rank fusion in hybrid search