How Many Searches Hide Behind One AI Prompt?

When a buyer types one question into Google's AI Mode, the engine does not run one search. It runs about nine.

A new study from Nectiv ran roughly 9,000 prompts through the Gemini 3 API and pulled the hidden sub-queries each one generated. The average prompt produced 9.06 fan-out queries. The most complex single prompt produced 28. For the software vertical, the average was 11.7.

So the answer to the question in the title is nine, and if you sell software it is closer to twelve.

One prompt is not one search. It is nine.

That gap is the reason a page can rank in Google's top 10 and still never get cited in the AI answer. The engine is not searching your keyword. It is searching the cloud of related questions around it. This post breaks down what the fan-out data shows, why it breaks the way most teams measure AI visibility, and the practitioner sequence for getting cited across the full fan-out set.

What the fan-out data actually shows

Query fan-out is the step where the model takes one prompt, splits it into subtopics, and fires a batch of separate searches at the web before it writes an answer. Google describes the mechanic in its own AI Mode announcement. Here are the five findings that matter, each one a separate number you can take into a planning conversation.

Finding #1: Google runs about 9 searches per prompt

Across the full Nectiv sample, the average prompt fanned out into 9.06 distinct queries. This is the headline figure and it holds across roughly 9,000 prompts in multiple verticals. One visible question becomes nine invisible ones, and the AI reads sources for each before it composes the response your buyer sees.

Finding #2: For software, it is closer to twelve

Fan-out volume varies by category. Local services averaged 3.79 fan-outs per prompt. Software averaged 11.7, the highest of any vertical measured. B2B SaaS buyers ask the most decomposable questions, so the engine does the most digging. If your category is software, your real search surface is roughly three times wider than a local business owner's.

Finding #3: Fan-out queries are short and specific

The average fan-out query was 6.7 words, and 77% of them ran 5 to 8 words. The longest measured 19 words. These are not the long, context-loaded prompts the buyer typed. They are tight, specific sub-questions the model wrote itself to retrieve clean facts.

Finding #4: Most fan-outs are comparison, review, and freshness queries

The most frequent terms inside fan-out queries were a recency marker like "2024/2025" at 6.26%, then "reviews" at 2.14%, "vs." at 1.41%, "free" at 1.05%, and "top" at 1.05%. The engine is systematically hunting for current comparisons, social proof, and pricing signals, not brand slogans.

Finding #5: Fan-out sets are unstable

Run the same prompt twice and the fan-out set shifts. Independent analysis puts the share of fan-out queries that repeat across runs at roughly a quarter. That means a single snapshot of "which queries the AI ran" is a sample, not a fixed list. Coverage has to be measured as a moving target.

You optimize the question. The AI searches the answers around it.

Why this breaks the way most teams track AI visibility

Most AI visibility programs were built on a one-prompt-equals-one-query mental model carried over from SEO. The fan-out data breaks that model in four specific ways.

Reason #1: Your 100 tracked prompts are really 900 searches

If you monitor 100 priority prompts, the engines behind them just ran on the order of 900 actual searches. Your tracking sees the 100 you chose. It misses the 800 the model wrote on its own. The denominator you are scoring against is roughly a tenth of the real one.

Reason #2: You optimize the headline query, the AI searches the sub-queries

A team builds one page for "best CRM for B2B," ranks it, and assumes the job is done. The engine answering that prompt is searching for "CRM pricing for 50 person company," "HubSpot vs Salesforce 2026," and "CRM SOC 2 compliance." If your content does not answer those, you are absent from the searches that actually feed the answer.

Reason #3: Ranking in the top 10 no longer means getting cited

Surfer SEO analyzed 173,902 URLs and found that 67.82% of pages cited in AI Overviews did not rank in Google's top 10 for the original keyword. The citation is being earned on the fan-out queries, not the head term. We covered the broader split in why Google rankings no longer predict AI citations.

Reason #4: A single snapshot misreads unstable coverage

Because only about a quarter of fan-out queries repeat between runs, a one-time check tells you almost nothing durable. You can look cited on Monday and absent on Thursday without changing a word. Coverage has to be sampled repeatedly, the way you would measure anything that moves.

If you track 100 prompts, the AI just ran roughly 900 searches you never logged.

The structural fix is the same finding Surfer landed on. Pages that ranked for the fan-out queries were 161% more likely to be cited in AI Overviews than pages ranking only for the main query, with a 0.77 correlation between fan-out coverage and citation odds. Coverage of the question cloud, not the keyword, is what gets pulled into the answer.

What your SEO targets versus what fan-out searches

The mismatch is easiest to see side by side. The left column is what most B2B content is still built for. The right column is what the engine actually queries.

What traditional SEO targets	What query fan-out actually searches
"best project management software"	"project management tools for remote teams 2026"
"[product] features"	"[product] vs [competitor] pricing"
"[category] platform"	"[category] reviews for mid-market SaaS"
"[product] for enterprise"	"[product] SOC 2 and SSO support"
One page, one keyword	Nine sub-questions, nine retrieval events

Each cell on the right is a self-contained question with its own answer. The engine retrieves a source for each one independently. A page that only answers the left column is invisible to most of the searches that build the response.

How to optimize for query fan-out

The findings above explain the gap. The sequence below is what we run for Cite clients to close it. Each step is independently runnable, and the order matters because measurement comes before rewriting.

Step 1: Expand each tracked prompt into its real fan-out set

Take your top 50 buyer prompts and pull the sub-queries each one generates, either through the Gemini API the way Nectiv did or through a tool that surfaces fan-outs. Expect roughly nine per prompt, more if you sell software. This turns a 50-prompt tracking list into a 450-query coverage map, which is the real surface you are competing on. The prompt-selection groundwork is in how to select prompts for LLM tracking.

Step 2: Score your citation coverage across the full set, not the head term

For each fan-out query, check whether your brand is cited and which domain earned the mention. Score coverage as the share of fan-out queries you appear in, not the share of headline prompts. Most teams discover their real coverage is far lower than their head-term number suggested, because the head term was the one query they optimized.

Step 3: Rewrite pages to answer the comparison, review, and pricing sub-queries

The data shows fan-outs lean on "vs," "reviews," "pricing," and recency. Add a direct 40 to 60 word answer for each high-frequency sub-query under a heading that mirrors it, plus a comparison table on any "vs" or "alternatives" page. Structuring content as extractable passages is the mechanic we cover in passages beat pages.

Step 4: Cover the fan-out cloud across multiple pages, not one mega-page

Nine sub-questions rarely fit cleanly on one URL. Build a small cluster where each page owns a few related fan-out queries: a comparison page, a pricing page, a use-case page. Surfer's data showed that ranking for more fan-out queries raised citation odds, so spread coverage rather than stuffing it. The on-page side is detailed in Google AI Mode optimization.

Step 5: Re-sample coverage on a fixed cadence

Because fan-out sets shift between runs and refresh every few weeks, measure on a schedule rather than once. Re-pull the fan-out sets and re-score coverage every 30 days, and treat any sub-query cluster where you drop out as a rebuild target. The absorption side of this is in how to measure AI citation absorption.

Ranking for the headline keyword is table stakes. Ranking for the fan-out set is the citation.

The findings sit inside a larger pattern Cite tracks across engines. ChatGPT's fan-outs roughly doubled in word length over four months, and Grok runs the most hidden searches of any engine. The number changes by surface, but the rule does not: the engine searches a cloud of questions, and you have to be in the cloud.

FAQ

What is query fan-out in AI search?

Query fan-out is the process where an AI search engine takes one user prompt, breaks it into subtopics, and runs multiple separate searches at once before composing an answer. Google uses the technique in AI Mode and AI Overviews. Nectiv's study found the average prompt generates about 9 fan-out queries, and the Search Engine Land guide walks through the mechanic in detail.

How many sub-queries does one AI prompt trigger?

Across roughly 9,000 prompts, Nectiv measured an average of 9.06 fan-out queries per prompt, with a maximum of 28. The number varies by industry: local services averaged 3.79 while software averaged 11.7. Other large-scale analyses land in the same 8 to 12 range for typical Google and ChatGPT queries.

Why does my page rank in Google but not get cited by AI?

Because the AI is not searching your keyword. It is searching the fan-out queries around it. Surfer SEO found 67.82% of AI Overview citations came from pages outside Google's top 10 for the original term. Pages that ranked for the fan-out queries were 161% more likely to be cited, so coverage of the sub-questions matters more than head-term ranking.

Does optimizing for query fan-out replace SEO?

No. Strong organic ranking still helps, and the Profound AEO Playbook notes that traditional SEO signals remain a factor in citation. Fan-out optimization adds a layer: instead of targeting one keyword per page, you map the cloud of sub-questions the engine actually runs and make sure your content answers them.

How often should I re-measure fan-out coverage?

Every 30 days is a reasonable baseline, because only about a quarter of fan-out queries repeat between runs and citation pools refresh on a multi-week cycle. A single snapshot is a sample, not a stable list, so coverage should be tracked as a moving metric rather than a one-time audit.

What to do this week

Take the three buyer prompts that most often precede a deal in your category. Run each one in Google's AI Mode and watch what it actually searches and cites. If the answer leans on comparison pages, review sites, and pricing breakdowns you do not have, you have just seen your fan-out gap in real time.

The Nectiv study did not change how AI search works. It measured something that was already true: one prompt was never one search. Teams that map the full fan-out set and score coverage against it, instead of against a list of head terms, are the ones who find out where they are actually losing the citation, and fix it before the next refresh cycle.