When someone asks Google's AI Mode a question, Google does not run that question. It breaks it into subtopics and fires a batch of searches on the user's behalf — an average of 10.7 per prompt in Seer Interactive's study of 501 Gemini 3 prompts, and 11.7 for software topics in Nectiv's larger sample of roughly 9,000 prompts. Here is the part that breaks the way most of us plan posts: 95% of the queries Gemini generated had zero global search volume. They will never appear in your keyword tool, at any volume threshold, ever. So query fan-out keyword research starts one level up from the volume column — you choose a topic, map the eight-to-fifteen sub-questions it fans into, and then decide honestly which of them a single post can answer in full.
That is a planning change, not a trick. But if you are still sorting a keyword tool by volume and assigning one head term per post, you are optimising for the query a human types — while the thing that actually retrieves your page is a machine typing ten queries you never saw.
Google described the mechanism itself. In its own words, "AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf" — and Deep Search does the same thing "at an expanded scale," issuing hundreds of searches for one prompt. Search Central confirms this is not an AI Mode exclusive: both AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response," which is how Google says it can show a wider and more diverse set of links.
The sub-queries are not random rewordings. Search Engine Land traces fan-out to patent US11663201B2, which Google calls "query variant generation," and catalogues eight sub-query types: equivalent, follow-up, generalization, specification, canonicalization, translation, entailment and clarification. A separate "Thematic Search" patent, filed December 2024, describes generating narrower sub-themes and cascading into sub-sub-themes — expand outward, then drill down.
What happens to the results matters as much as the expansion. Mike King's teardown of AI Mode documents retrieval at the passage level over dense vector embeddings rather than at page level: "every query, subquery, document, and passage is converted into a vector embedding," and the model fans out "across this hidden web of synthetic queries, scanning not just for facts, but for ideas that can complete a 'reasoning chain.'" So a single section of your post, not the post as a whole, is the unit that gets retrieved. Hold that thought — it is what makes one thorough post a legitimate answer to a fan-out rather than a compromise. And the audience is real: Sundar Pichai told investors in July 2026 that AI Mode had "surpassed 1 billion monthly active users", while Search Senior Engineering Director Dounia Berrada sums up the mechanism as "basically doing a dozen searches for you in the time it takes to do one."
You will see "8–12 sub-queries" repeated everywhere as settled fact. Treat it as folklore until someone publishes the sample. The measured studies are more useful anyway, and they broadly agree with each other.
Seer tracked 501 prompts through Gemini 3 with forced grounding and found 10.7 fan-out queries on average, ranging from 3 to 28 — a 78% jump from Gemini 2.5, which averaged 6.01. Nectiv's larger run, roughly 9,000 prompts through the Gemini 3 API producing over 60,000 fan-out queries, landed at 9.06, with 5 the most common count. That distribution is the number to plan around: 59% of prompts fired between 5 and 11 sub-queries, 24% fired between 12 and 19, and the maximum was 28.
Sixteen is real, then, but it is upper-quartile behaviour rather than the average — roughly one prompt in four fans out that far. Vertical matters too, and it matters in your favour if you sell software: Nectiv clocked software topics at 11.7 sub-queries per prompt, travel at 10.8, careers at 9.8 and local at 3.79. Someone researching a tool triggers three times the machine-side research that someone looking for a nearby plumber does.
Engines differ as much as verticals. EMGI's July 2026 pilot on SaaS buying prompts recorded ChatGPT averaging 6.3 sub-queries per searching prompt — well under Gemini 3's 10.7 — and found the bigger variable is whether the model searches at all. On those prompts Perplexity searched 100% of the time, Claude 81%, ChatGPT 42% and Gemini's assistant just 6%, answering 15 of 16 prompts from memory. When a model answers from memory, nothing you publish can influence that answer; only what it already absorbed about you counts. That is the honest boundary around everything below.
Two findings from the Seer data do the real damage. The first is the 95% with zero global search volume. The second is quieter and worse: there was "only 1% overlap across the full fan-out dataset, so almost every query Gemini generated was unique," in the words of Seer R&D lead Nick Haigler. Even if you harvested ten thousand real fan-out queries, you would not be holding a keyword list. You would be holding samples from a distribution that never repeats exactly.
None of this is new, which should be reassuring. Ahrefs finds keywords with fewer than ten monthly searches "account for almost 93% of our U.S. keyword database" — 2.3 billion of them — with roughly 15% of daily Google searches never having been searched before. And one page serving hundreds of queries was already normal: Ahrefs' study of three million searches found the average #1 ranking page also ranks in the top 10 for nearly 1,000 other keywords, median around 400. That study is from 2017. "One post, one keyword" was a bookkeeping convention long before AI Mode existed.
Human behaviour drifted the same way. Google Search VP Robby Stein said AI Mode testing showed roughly a two-to-three-fold increase in query length, his example being "restaurants to go to in Nashville if one friend has an allergy and we have dogs" instead of "things to do in Nashville." Longer human questions, expanded into more machine questions, resolved against passages rather than pages — and your volume column can see only the first layer of that.
Here is the good news, and it is the whole method. You cannot target individual fan-out strings. But the shapes those strings take are strikingly consistent across four independent datasets and two model families.
Seer and Nectiv both measured an average fan-out query length of 6.7 words, with Nectiv finding 77% between five and eight words. Nectiv's most common n-grams were year references ("2024/2025," 6.26%), "reviews" (2.14%), "vs." (1.41%), "free" (1.05%) and "top" (1.05%); Seer found 21.3% of queries contained a year and 26.4% named a brand. Centium's much larger sample — 322,485 fan-out searches across 130+ brands through ChatGPT and Gemini over three months in 2026 — reports that "four in five of the prompts AI researches (80%) trigger at least one search for the best or top option," over a third trigger a review search, about 15% trigger a guide search, and nearly half include a year-specific search. On SaaS buying prompts specifically, EMGI's 127 captured sub-queries were 86% year-stamped, 69% brand-named and 51% about pricing.
One conclusion, then: the machine researching your category asks priced, dated, brand-named, compared and reviewed questions in five to eight words. That is not a keyword list. It is an outline.
| Fan-out shape | Evidence | What it wants on the page |
|---|---|---|
| Priced ("X pricing 2026") | 51% of SaaS sub-queries (EMGI) | A real number or named range, and what changes it |
| Year-stamped | 21.3% (Seer), 86% of SaaS sub-queries (EMGI) | A visible date and current figures, not "recently" |
| Brand-named | 26.4% (Seer), 69% (EMGI) | Named competitors and integrations, described fairly |
| Comparative ("vs") | 1.41% of n-grams (Nectiv) | A side-by-side with an honest "choose them if…" |
| Best / top | 80% of researched prompts (Centium) | A shortlist with stated selection criteria |
| Reviews | Over a third of prompts (Centium) | Attributed third-party evidence, not self-praise |
| How-to / guide | ~15% of prompts (Centium) | Numbered steps that stand alone as a passage |
EMGI's captured example makes it concrete. Run "best CRM for small B2B sales team" through GPT-5.2 with live search and it generated eight sub-queries, including "HubSpot Sales Hub pricing 2026" and "Pipedrive pricing 2026 plans." Nobody types those. A machine does, on its way to answering the question somebody did type.
Cyrus Shepard's five-step fan-out framework is the closest thing the field has to a practitioner standard, and it adapts cleanly to a one-person content operation.
1. Start from a query you already have impressions for. Not a volume winner — an existing foothold in Search Console. Fan-out rewards pages that are already retrievable, and Shepard's blunt version is worth taping to the wall: "You really want to create content worthy of ranking in the top 10-20 of Google's search results."
2. Harvest observed fan-outs before simulated ones. Every fan-out generator is a model guessing at what Google might ask; two sources give you queries a model actually issued. Bing Webmaster Tools' AI Performance report, in public preview since February 10, 2026, includes a Grounding Queries metric Microsoft defines as "the key phrases the AI used when retrieving content that was referenced in AI-generated answers" — free, and about your own pages. Second, the Gemini API returns the searches it ran in a google_search_call object, and grounding gives you 5,000 free search requests a month on Gemini 3.x models, then $14 per 1,000 — at roughly ten searches per prompt, about 500 prompts a month of observed data.
3. Use simulators as a fallback, knowingly. Qforia, DEJAN's tool, Wellows' generator, Otterly's fan-out analysis and Locomotive's coverage scorer all appear in Search Engine Land's tools roundup, and they are genuinely useful for structure — Qforia classifies its output by query type, user intent and recommended format, from tables to checklists. Just log the results as hypothesis, not observation.
4. Cluster into eight to fifteen sub-questions, then cut. Group the harvested strings by the shapes in the table above and de-duplicate ruthlessly. Then apply the honesty test: which of these can this post answer in full, with a number or a procedure, from evidence you actually have? A sub-question you can only wave at belongs to a different asset, or nowhere.
5. Write each surviving sub-question as a self-contained section. Since retrieval happens at passage level, every section needs its claim, its number and its scope in the same place — a paragraph that still makes sense with the rest of the post deleted. iPullRank's worked example of "What's the best way to save for retirement?" fans into "what are the different types of retirement accounts?", "how much money should I save for retirement?" and "common retirement savings mistakes to avoid" — one wants a table, one a calculator, one a checklist. Match the format to the sub-question instead of flattening all of them into prose; our guide to the on-page patterns that earn AI citations goes deeper on shaping those paragraphs.
Steps 4 and 5 are where a solo operator runs out of hours. Answering twelve sub-questions properly, each with its own sourced evidence, is a research job wearing a writing job's clothes — which is the shape of work Magic Share's drafting and verification pipeline was built around: 2,000–4,000-word drafts with every cited URL resolved before the draft is saved, a verification report attached, and the draft waiting for a human to decide which sub-questions the post is entitled to answer.
The obvious wrong turn is to read a twelve-query fan-out as a twelve-post content calendar. Three reasons not to.
First, the policy language. Google's spam policies define doorway abuse as "sites or pages … created to rank for specific, similar search queries" — a fair description of one thin post per sub-query. Our closer read of what Google's scaled content abuse policy actually bans covers where that line sits; the short version is that value per page is the test, and method is irrelevant.
Second, the correlation data. Ahrefs' study across 75,000 brands found the number of site pages among the weakest correlates of AI visibility, at roughly 0.194, while YouTube mentions (~0.737) and branded web mentions (0.656–0.709) sat far higher — with the authors' own correlation-isn't-causation caveat attached. Publishing more pages is the lever with the least visible relationship to the outcome you want.
Third, what worked in practice was consolidation. Resource Guru, a SaaS in resource management, spent Q1 2026 refreshing existing posts across three clusters — project time tracking, agency software and resource capacity planning, 44 keywords in total — updating angle and UX and adding expert insight, entity markup, FAQ blocks and schema. Share of voice on those keywords went from 1.82% to 9.84%, ten of them earned AI Overview citations, and organic traffic moved from 17,614 to 22,546 monthly visits. It is an agency-published case study, so treat it as directional — but the direction matches Shepard's advice to add fan-out coverage to pages that already exist.
The rule that falls out: sixteen sub-queries is an outline, not a calendar. Split only when a sub-question needs a genuinely different format or serves a different stage of intent — a pricing page, a calculator, a migration guide. If two candidate posts would answer overlapping sub-questions, you have invented a cannibalisation problem instead of solving a coverage one. One seed, many shoots; you don't plant sixteen seedlings in one pot.
Not every sub-query is winnable, and pretending otherwise wastes the scarcest resource you have. The "best [category] software" branch is largely spoken for. In a June 2026 analysis of 1,259 AI Overview citations across 100 commercial "best software" searches in 20 B2B SaaS categories, third-party best-of lists took 63% of citations while the recommended product's own site took just 12% — an agency study, so directional, but consistent with Centium finding Gemini citing best/top list pages in 44.6% of its cited URLs. Semrush's cross-platform work adds that Reddit dominated across AI Mode, AI Overviews, ChatGPT and Perplexity. Writing your own "best CRM tools" listicle to win that branch means beating every review site on its home turf.
The branches vendors win are the ones only vendors can answer. EMGI found 56% of ChatGPT's citations went to vendor-owned pages, mostly pricing and comparison content — which lines up with 51% of its captured sub-queries asking about pricing. So the instruction is specific: spend your effort on published pricing with the conditions that change it, honest head-to-head comparisons, integration and migration questions, limits and specs, and implementation timelines. Those sub-queries fan out constantly, and your own documentation is the most authoritative surface on the web for them.
There is a structural reason this is a better bet than it used to be. Ahrefs found that 38% of AI Overview citations come from pages ranking in the top 10 — across 863,000 keyword SERPs and 4 million AIO URLs in March 2026 — down from about 76% in their July 2025 study of 1.9 million citations, a change they attribute to AI Overviews "relying less on the direct search results and more on the sources showing up in fan out query SERPs." Semrush's 5,000-keyword comparison points the same way: AI Mode has the loosest relationship with Google's top 10, at roughly 54% domain and 35% URL overlap. Ranking first for the head term matters less than it did; being the best available answer to one specific sub-query matters more. For a small site, that arithmetic is a gift.
Measurement is the weak link, and it is asymmetric. Google folds AI feature performance into ordinary reporting — pages appearing in AI features are "included in the overall search traffic in Search Console," reported within the "Web" search type, with no fan-out breakdown. Microsoft hands you grounding queries for free, though Bing's report has real limits: Search Engine Land notes it tracks citation frequency only, not prominence or traffic attribution, and Microsoft expanded it on June 16, 2026 with intent labels, topic groupings and citation share. For a solo marketer, the smaller engine is currently the better instrument. Use it as a proxy: if your page is retrieved for six of the twelve sub-questions you mapped, the missing six are your next editing session.
Then the question underneath all of this: is any of it worth the hours? Pew's tracking of 900 U.S. adults across 68,879 searches in March 2025 found users clicked a traditional result in 8% of visits to pages with an AI summary versus 15% without, and clicked links inside the summary in just 1% of visits. Fewer clicks, plainly. Semrush argues the survivors are worth more — the average AI search visitor being 4.4 times as valuable as a traditional organic visit by conversion rate across 500+ topics — but that figure is contested, with MarTech noting Amsive found no significant difference in conversion. Don't build a business case on 4.4x. Our take on what a blog is still for when most searches end without a click sets out the version we would defend: the job shifted from capturing traffic to being the source the answer is assembled from.
The last word belongs to Google's Search Liaison, Danny Sullivan: "SEO for AI is still SEO," and if you optimise narrowly for one AI system, "you risk permanent catch-up as those systems evolve." Search Central says something similar — there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." Read alongside the fan-out data, that is not a contradiction. Fan-out doesn't change what wins; it changes how many distinct questions one page has to satisfy to keep being found. Mapping durable question shapes instead of chasing this month's model-specific strings is exactly the non-narrow approach Sullivan is recommending.
How many sub-queries does one search actually become? Measured averages cluster between 9 and 11. Seer recorded 10.7 per prompt across 501 Gemini 3 prompts (range 3–28); Nectiv recorded 9.06 across roughly 9,000 prompts, with 59% firing 5–11 sub-queries and 24% firing 12–19. Software topics run higher, at 11.7. Sixteen happens, in that upper quartile, but it isn't typical — and the widely-quoted "8–12" range circulates without a published methodology.
Should I still use keyword volume at all? Yes, for the entry query. Volume tells you what humans type, which is what your title and framing target. It cannot tell you anything about the sub-queries a model generates on the way to answering, since 95% of those had zero global search volume in Seer's data. Use volume to pick the door; use fan-out shapes to furnish the rooms.
Are fan-out simulators trustworthy?
They are useful and they are guesses. Tools like Qforia and the DEJAN and Wellows generators produce plausible sub-queries from a model, not Google's actual expansion. For observed data, Bing Webmaster Tools' grounding queries show real phrases the AI used to retrieve your cited content, and the Gemini API returns the queries it issued in its google_search_call object.
Is it better to write one thorough post or several short ones per fan-out? Consolidate by default. Google's doorway-abuse wording targets pages created to rank for "specific, similar search queries," Ahrefs found page count among the weakest correlates of AI visibility (~0.194), and the practitioner advice is to add fan-out coverage to existing pages. Split only when a sub-question needs a different format — a pricing page, a calculator, a migration guide.
One search became ten, most of the ten are invisible to your tools, and the shapes they take — priced, dated, brand-named, compared, reviewed — are stable enough to plan around. Pick your entry query the way you always did, then spend the planning time mapping the eight-to-fifteen sub-questions behind it and cutting the ones this post cannot honestly answer. That map is your keyword research now.
If the bottleneck is doing that research every week rather than knowing how, that is the gap Magic Share was built for: the agent studies your niche, drafts the post with every cited link resolved and a verification report attached, then leaves it as a draft until you approve it. Your first three posts are free — no credit card, just a spot in line.
Magic Share researches, writes and fact-checks posts like this for any site — point it at your URL and review your first draft today.
Plant your first post