Your page is judged on a query nobody typed

Query fan-out is the step where an AI engine rewrites a typed question into its own set of search strings, and those strings, not the original question, decide which pages get retrieved. Google describes the mechanic in its own words: AI Mode "uses a 'query fan-out' technique, issuing multiple related searches concurrently across subtopics and multiple data sources" (blog.google).

Most fan-out advice treats this as one behavior with one number attached. The 2026 measurements say it is two independent variables: whether the engine searches the live web at all, and how far it spreads when it does. They do not move together, and on some engines they move in opposite directions.

#The number in circulation describes one engine

The figure most people quote is 9 to 11 fan-out queries per prompt, from Seer Interactive and Nectiv, with 59% of prompts triggering 5 to 11 searches, 24% triggering 12 to 19, and outliers reaching 28 (Ahrefs).

Measured across chat assistants at scale, the averages are much lower. Peec AI analyzed 5 million query fanouts collected between April 1 and April 21, 2026 and found Perplexity averaging 1.4 fanouts per query, ChatGPT 2.1, and Grok 6.8, more than triple ChatGPT (Peec AI).

Both sets of numbers can be correct. They describe different populations: exploratory prompts on a Google surface built for comparison and reasoning, versus everything people type into three general assistants. The practical consequence is that "plan for a dozen sub-queries" is advice for one surface, not a universal constant. Planning your content depth against the wrong engine wastes the effort.

#Firing and fanning are different decisions

The second variable gets ignored because it is invisible in fan-out counts. An engine that never searches has no fan-out at all, and no amount of on-page work will surface you inside that answer.

EMGI ran 48 SaaS buying prompts through ChatGPT in July 2026 and a 16-prompt subset through Gemini, Claude and Perplexity. ChatGPT triggered a live web search on 42% of buying prompts, 20 of 48. Perplexity triggered one on 100% of them, 16 of 16 (EMGI).

Line that up with the breadth data and the engines separate cleanly. Perplexity nearly always goes to the web but barely expands the query, so the prompt you track is close to the string that actually runs. ChatGPT goes to the web selectively and expands when it does, so the retrieval happens in sub-queries you never see. Grok expands hardest of the three and leans on explicit source targeting: site operators appear in 18.3% of Grok chats, with Reddit in 10.5% of all chats (Peec AI).

Three different failure modes follow. On Perplexity you lose by not ranking for the prompt. On ChatGPT you lose by not covering the adjacent sub-queries. On Grok you lose by not being present on the handful of sites it scopes to.

#What the engine actually types

The rewrites are not neutral rephrasings. They add commercial and temporal modifiers with striking consistency.

At Peec AI's scale, "best" is ChatGPT's most common injection, appearing in 24.3% of fanouts for advice-style questions, with top, comparison, reviews, tools, software and features close behind. The current year gets added to 5.44% of prompts (Peec AI).

On buying-intent prompts the skew is far stronger. Of the 127 sub-queries EMGI captured, 86% contained a year reference, 69% included a specific brand name, and 51% asked about pricing (EMGI).

That is a precise description of the page that wins a buying-journey retrieval: dated, named, and priced. Not a general capability page. A page that survives a search for your brand plus a competitor plus a year plus the word pricing.

There is an encouraging finding attached. 56% of ChatGPT's citations in that study pointed to vendor-owned pages rather than third-party roundups, with pricing and comparison pages dominating the vendor-owned share (EMGI). On commercial queries, your own pages are eligible. The sample is small, 48 prompts in one vertical, so treat the direction as real and the exact share as provisional.

#Keyword tools cannot see any of this

Fan-out strings are synthetic. They exist inside one reasoning chain for a few milliseconds. More than 95% of them receive no recurring searches at all (Ahrefs), which means the queries deciding your retrieval have no volume, no difficulty score and no entry in any keyword database.

The depth ceiling is higher than most planning assumes, too. One documented case had ChatGPT Deep Research run 420 searches for the query "buy red phone case" (Ahrefs). Agentic modes are a different regime from chat, and the same page can be evaluated against hundreds of strings.

#How we would work this

Four changes follow from the data above.

Stop treating a fan-out average as a target. Establish which engines your buyers actually use, then plan depth per engine. A Perplexity-heavy audience rewards ranking for the prompt itself. A ChatGPT-heavy audience rewards covering the neighbours.

Capture real fan-outs instead of guessing them. The sub-queries are observable if you record what the engine retrieves rather than only what it answers. That is the input worth collecting, and it is the one most teams skip.

Write the modifier into the page. If year, brand and pricing appear in the majority of buying-intent sub-queries, then a comparison page with a visible date, named competitors and real numbers is not keyword stuffing, it is matching the string that runs.

Track search-trigger rate as its own metric. A prompt where the engine answers from memory is not a prompt you lost on merit. It is a prompt where retrieval never happened, and the fix is brand presence in training-adjacent sources, not another rewrite of the page.

The older framing still holds: cover the questions around the question, as we wrote in query fan-out: answer the hidden queries. What the 2026 data adds is that the shape of that coverage is engine-specific, and that half the battle is fought before any fan-out happens at all.