Query Fan-Out: What the AI Actually Searches When Your Buyer Asks a Question

Original Research · LLM SEO · July 2026

When a buyer asks ChatGPT for “the best CRM for a small B2B sales team”, the answer they read was never one search. Under the hood, the model quietly ran eight searches of its own, checking pricing pages, comparing named vendors, appending “2026” to everything, and it built the answer from whatever those hidden searches returned. Google calls this query fan-out. Almost nobody has published what those hidden searches actually look like for B2B SaaS buying questions. So we captured them.

I’m Matt, founder of EMGI, an LLM SEO agency for SaaS. This is a pilot study: 48 real SaaS buying prompts run through GPT-5.2 with live web search, and a 16-prompt subset re-run across Gemini, Claude and Perplexity so we could compare engines on identical questions. Small sample, honestly framed, and it still surfaced three findings that change how you should think about AI visibility.

One question in, eight searches out

Query fan-out is Google’s own term. Announcing AI Mode at I/O 2025, Google’s Head of Search described it as “breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf” (Google, May 2025), and Google’s Search Central documentation now defines it formally.

The published numbers are consistent. Seer Interactive measured 10.7 fan-out queries per prompt on Gemini 3 across 501 prompts. Nectiv’s analysis of 60,000+ captured fan-outs found an average of 9.06, with software the highest fan-out vertical of all at 11.7. Around 95% of these sub-queries have no measurable search volume, which is why your rank tracker has never heard of them, and Ahrefs reports ChatGPT’s Deep Research running 420 searches for a single task.

The engine then merges the results of all those sub-queries, using techniques like reciprocal rank fusion, and sources that appear across several result lists accumulate score. That one mechanical detail is why appearing in many of the hidden searches beats ranking first for the visible one. We’ll go deeper on that mechanism in a follow-up piece.

How we captured real query fan-out data for SaaS buying prompts

The public research reports aggregates: how many fan-outs, how long, how often a year appears. What nobody had published is the actual sub-queries behind B2B SaaS buying prompts, so that is what we collected.

We wrote 48 prompts across six SaaS categories (CRM, project management, HR, marketing automation, dev tools and customer support), split deliberately across the funnel: “what is a kanban board” at the top, “how to choose a CRM” in the middle, “best X” and “[vendor] alternatives” at the bottom. Each ran through GPT-5.2 with live web search via the DataForSEO LLM API in July 2026, logging the fan-out queries, every cited source, and the brands each answer recommended. A 16-prompt subset then re-ran identically on Gemini 3.1 Pro, Claude Opus 4.8 and Perplexity Sonar Pro.

Each prompt ran once (single-shot, with a retry only on API failure), and we logged four fields per run: whether the engine searched at all, the fan-out queries it generated, every source it cited, and the brands its answer recommended. The funnel split was four bottom-of-funnel, two middle and two top-of-funnel prompts per category, mirroring the structure of our earlier Reddit research so the two studies can be read together.

The full raw dataset, all 96 runs with every captured sub-query and cited domain, is downloadable here: the query fan-out dataset (CSV). If you cite this study, cite it as “EMGI query fan-out capture, July 2026”.

Limitations, stated plainly: this is a pilot, not a population study. API behaviour isn’t guaranteed to match each engine’s consumer app exactly, single runs of non-deterministic systems will vary, and Perplexity’s API exposes citations but not its internal sub-queries. Treat the percentages as strong directional signals. The structural findings, what the sub-queries look like and how the engines differ, are the point.

Fan-out queries, in plain English

Think of the AI as a diligent assistant. Ask a good assistant to “find a restaurant for Friday” and they do not run one search. They check opening hours, skim the menu, compare prices and read a couple of reviews, separately. Fan-out queries are exactly that, written by a machine: the model reads your question, decides what it would need to know to answer it well, and fires each of those checks as its own mini-search.

Three things make them strange. First, they are invisible: no user ever sees them, and the only way to know they exist is to capture them at the API level, which is what this study does. Second, they are written in machine shorthand rather than human phrasing (“Pipedrive pricing 2026 plans”), so they rarely look like anything in your keyword tool. Third, because the model writes them fresh each time, almost none of them carry search volume. Your visibility in that hidden layer is decided by content and mentions you built for other reasons.

And here is the merging step, in plain terms. When the results come back from all those mini-searches, the engine combines them like a panel of judges: each mini-search ranks the pages it found, and a page collects points every time it appears near the top of any judge’s list. A page that shows up on five of eight lists beats a page that tops just one. That, in a paragraph, is what reciprocal rank fusion does, and it is why breadth of presence now beats a single high ranking. Google has not confirmed its exact method, but the patents and standard retrieval practice point this way.

Finding 1: does the AI even search, or answer from memory?

We expected to study which sources get cited. The bigger finding arrived first: on most of our prompts, several engines never went to the web at all.

Gemini was the extreme case. Its reasoning traces said things like “I don’t need to dig through the web; this is all standard knowledge”, and it answered fifteen of sixteen prompts entirely from memory, including “best HR software for small business”. ChatGPT split cleanly by intent. Every “best X for Y” and “alternatives” prompt triggered a search. Definitional questions, “how to choose” questions and, curiously, almost every two-way “X vs Y” comparison were answered from memory (the three-way Asana vs Monday vs ClickUp comparison did search; the two-way ones did not). Perplexity searched every single time, and Claude searched on 13 of 16.

Here is why that matters commercially. For the same prompt, “HubSpot alternatives for mid-market companies”, compare what the buyer gets:

Gemini’s from-memory list still recommends Pardot, a brand name Salesforce retired years ago. That is what a memory answer looks like: frozen at training time, uncorrectable by any page you publish. When the model answers from memory, there’s no page you can rank, no listicle you can appear in, and no citation to win. Your only lever on those prompts is being so consistently present across the web that the next training run knows who you are, which is a brand-mention and authority play, not a content play.

Finding 2: query fan-outs speak in brand names, years and prices

When ChatGPT did search, the sub-queries were not paraphrases of the question. They were a buyer’s due-diligence checklist, written by the machine. Across all 127 captured fan-out queries:

Here is the full fan-out set for one prompt, “what is the best CRM for a small B2B sales team?”, exactly as the model ran it:

best CRM for small B2B sales team 2026 HubSpot Pipedrive Zoho Freshsales pricing comparison
HubSpot Sales Hub pricing 2026 Starter Professional
Pipedrive pricing 2026 plans
Zoho CRM pricing 2026 standard professional enterprise
Close CRM pricing 2026 per seat startup small sales team
Freshsales (Freshworks CRM) pricing 2026 Growth Pro Enterprise
Zoho CRM pricing official 2026 per user Standard Professional
Salesforce Starter Suite pricing 2026 sales cloud small business

Read that list again as a SaaS founder. The model chose which vendors to research before it searched. Seven of eight sub-queries name specific brands, and if your product isn’t part of the model’s vocabulary for the category, you were never in the running, no matter what you rank for. And because the sub-queries hunt for current pricing, a pricing page the model cannot read (JavaScript-rendered tables, prices in images) hands your answer to whichever third party lists your prices in plain text.

Finding 3: who does AI search actually cite?

Across ChatGPT’s citations, 56% went to vendor-owned pages, overwhelmingly pricing pages and the vendors’ own comparison content. ClickUp’s own “best project management software” listicle was cited on category queries. Gusto’s “BambooHR competitors” page was cited on a competitor query. The single most-cited third-party domain was G2, but far behind the vendors themselves. The lesson: your own comparison and alternatives pages aren’t just conversion assets, they’re citation assets.

The engines also have distinct personalities. Perplexity’s most-cited domain in our sample was YouTube, with 20 citations, followed by vendor comparison pages, Reddit on “alternatives” prompts, and a long tail of niche review blogs (Perplexity’s API does not expose its sub-queries, but it averaged 11.6 citations per answer, the widest source set of any engine). Gemini’s single search session cited mostly low-authority listicle sites. Claude leaned on smaller independent blogs. Optimising for “AI search” as one thing is already the wrong frame: each engine reads a different web.

What this means if you sell SaaS

• Get into the category vocabulary. The fan-outs name brands. Contextual brand mentions across the round-ups, comparisons and communities the engines read are what put you in that list. This matches what our 150-brand citation study found: topical authority predicts AI citations at r = 0.76, while raw traffic barely registers.
• Make your pricing machine-readable. Half the fan-outs hunt pricing. Plain-HTML pricing beats a beautiful JavaScript table that quietly hands your citation to an aggregator.
• Build your own comparison and alternatives pages. The engines cite vendors’ own comparison content constantly. It’s the most under-used citation asset in SaaS.
• One strong page per claim, not a page per fan-out. Google’s spam policy explicitly warns against generating pages to chase individual fan-out queries, and results dedupe by domain anyway. Depth beats sprawl.
• Spread across surfaces, deliberately. Perplexity reads YouTube and Reddit. ChatGPT reads pricing pages and G2. Gemini mostly reads its own memory. Our Reddit research and directory study cover two of those surfaces in depth; the point is the portfolio, not any single channel.

The key numbers, stated plainly

If you are citing this study, these are the headline statistics in one place (EMGI query fan-out capture, July 2026):

• 42% of SaaS buying prompts triggered a live web search on ChatGPT (20 of 48). Gemini searched on 6% of prompts (1 of 16), Claude on 81% (13 of 16), and Perplexity on 100% (16 of 16).
• ChatGPT generated 127 fan-out queries across 48 SaaS buying prompts, an average of 6.3 per searching prompt.
• 69% of ChatGPT’s fan-out queries contained a specific brand name.
• 86% of ChatGPT’s fan-out queries contained a year, and 51% asked about pricing.
• 56% of ChatGPT’s citations pointed to vendor-owned pages rather than third-party editorial.
• YouTube was Perplexity’s single most-cited domain, with 20 citations across 16 answers.
• Gemini’s from-memory recommendations included a brand name its vendor retired years ago.

Frequently asked questions

What is query fan-out?

Query fan-out is the technique AI search engines use to answer one question by silently running many related sub-queries, then merging the results into a single answer. The term comes from Google’s AI Mode announcement, and published studies measure roughly 9 to 11 sub-queries per prompt, with software queries fanning out hardest.

Do fan-out queries have search volume?

Almost never. Around 95% of captured fan-out queries (sometimes written “fanout queries”) have no recurring search volume, because the model writes them on the fly. That’s why keyword tools cannot see them and why topic coverage now matters more than chasing individual keywords. Seer Interactive found the same pattern in its Gemini research, where 26% of captured fan-outs contained a brand name; in our SaaS-specific capture, that figure rose to 69%.

Should I create a page for every fan-out query?

No. Google’s spam policies explicitly call out scaled page creation aimed at fan-out queries, and AI engines deduplicate results by domain, so twenty thin pages collapse into one. Build one genuinely strong page per real claim or comparison instead.

How does AI search pick which sources to cite?

In our capture, ChatGPT cited vendor-owned pricing and comparison pages most (56% of citations), with G2 the top third party. Perplexity leaned on YouTube, Reddit and review blogs. Sources that appear across several fan-out result lists accumulate score, so broad presence beats one strong page.

Does ranking for fan-out queries make you more likely to be cited?

The industry is genuinely split. Surfer’s analysis of 173,000 URLs found pages that also rank for fan-out queries are 161% more likely to be cited in Google AI Overviews, while Kevin Indig’s 815,000-pair study with AirOps found subtopic coverage added only around 4.6 percentage points. Our capture points to a middle reading: being named in the fan-out vocabulary matters most, then being retrievable for the sub-queries that follow.

How do I find the fan-out queries for my category?

The method in this study is repeatable in five steps: (1) write 15 to 30 prompts your real buyers would ask, across the funnel; (2) run each through an LLM API with web search enabled (we used the DataForSEO LLM API against GPT-5.2, Gemini, Claude and Perplexity); (3) log the fan-out queries, citations and named brands per run; (4) map which sub-queries and sources you currently appear in; (5) treat the gaps as your content and mention targets. We run this capture for clients as part of our audits.

This is the fourth instalment of our original research series, alongside the SaaS AI Citation Gap Report linked above, and it pairs with our tactical guide on getting your SaaS cited in ChatGPT. Next in the technical series: reciprocal rank fusion, the maths that decides which query fan-out results win.