Search volume measures the questions people type into a box. But AI answers are increasingly assembled from questions no one types: synthetic sub-queries the model writes for itself. Here's the shift, shown with a single prompt, and why that same prompt is also a content, paid and social brief.
Modern SEO has one founding instrument, and it's search volume. Keyword research is the ritual: pull the list, sort by monthly searches, chase the high-volume heads, and quietly delete the long tail that isn't “worth” a page. Every content calendar, every brief, every “is this term worth targeting?” decision runs through that one number. It has been the closest thing the discipline has to a measure of demand.
The number was never perfect, but it was honest about one thing: it counted queries humans actually typed. That assumption held for twenty years. Not so much anymore.
When someone asks an AI a question, the model usually doesn't search that question. It runs a process called query fan-out: it breaks the prompt into a set of related sub-queries, retrieves results for each one in parallel, then synthesises them into a single answer. Google coined the term for AI Mode and describes the underlying method in a patent as “query variant generation”, and the same retrieval pattern sits under AI Overviews, ChatGPT, Perplexity and Copilot.
The sub-queries aren't random. The model deliberately generates comparative variants (X vs Y), narrower angles, and, crucially, implicit and personalised queries. The unstated needs behind the question, tailored to who's asking. This is why “I'm a woman looking for running shoes” quietly becomes “best running shoes for women” before a single result is fetched. The user never typed it. The machine did.
And here is the fact the whole discipline has to sit with: these fan-out queries are synthetic. They are explicitly not long-tail keywords, which are real phrases real people have typed. A complex prompt can fan out into eight, twenty, or more of them. None of them have search volume, because no human searched them.
Theory is easy to wave away, so here’s a live one. We took a completely ordinary shopping prompt, “I need to buy a bag that fits as a carry-on luggage”, and ran it through Otterly.ai’s fan-out tool, which models how ChatGPT, AI Overviews and AI Mode expand a query. Then we checked every resulting sub-query against Ahrefs.
| Sub-query the prompt fanned out into | Type | Ahrefs monthly volume |
|---|---|---|
| best carry-on luggage for international flights | Best-of | 0–10 |
| best professional carry-on bags for business travellers | Best-of | 0–10 |
| top-rated affordable carry-on luggage for budget travellers | Best-of | 0–10 |
| hardshell vs softside carry-on luggage durability | Comparison | 0–10 |
| rolling carry-on suitcase vs travel backpack comparison | Comparison | 0–10 |
| airline carry-on size and regulation checks | Constraint | 0–10 |
Prompt run through Otterly.ai’s fan-out tool. Volumes measured in Ahrefs.
Look at what the model did on the user's behalf. It didn't just look for “carry-on bag”. It segmented the buyer, international flyer, business traveller, budget traveller, and generated a dedicated “best of” query for each. It reached for two head-to-head comparisons. And it ran compliance checks on airline size rules, because it correctly inferred the whole point of the purchase.
Every one of those buyer-intent queries, the three rankings and both comparisons, returns 0–10 monthly searches in Ahrefs. In a traditional keyword project, all five would have been invisible, or flagged as too small to bother with, or cut from the sheet to keep the focus on high-volume terms. They are exactly the rows we've been trained to delete, and they are exactly the rows the machine used to build its answer.
The problem isn't that search volume is wrong, it's that it's measuring the wrong layer. Volume counts human keystrokes, in a system that's no longer everyone's default. Fan-out queries are machine output, but they come after the same human input, just decomposed in a different setting. A tool that only sees the first can, by construction, never show you the second. So the more of the buyer journey the model absorbs into synthetic queries, the more demand slips below the waterline of every keyword tool you own.
And it isn't only machines producing queries volume can't see. People have started searching this way too. In May 2026 Google rebuilt its search box for the first time in over 25 years, a field that expands to take long, conversational questions, formalising a habit that had already taken hold as ChatGPT and Perplexity trained millions to ask in full sentences. But those human queries don't register either, for a different reason: they fragment.
A thousand people wanting the same thing phrase it a thousand slightly different ways, so no single string ever accumulates volume. The exact query might be typed once. The demand behind it is enormous. Volume is blind to the machine’s synthetic queries and to the scattered fragments of human ones alike. It undercounts both.
Two consequences follow. First, visibility stops being binary and becomes probabilistic. In classic search you either rank for a keyword or you don’t. In AI search you might not rank for the head term at all, yet get retrieved and cited across a dozen sub-queries that decide the answer.
Second, the “niche” term inverts. A phrase with near-zero volume used to mean near-zero opportunity. Now it can mean you’re one of the few credible sources in a data void, and being the source the model reaches for in a thinly covered space is worth far more than being the ninetieth page competing for a head term.
Notice the shape of what the carry-on prompt produced: rankings and comparisons. That's not a quirk of luggage. Comparative and multi-criteria decisions are precisely the prompts that fan out most aggressively, and the sub-queries they spawn are overwhelmingly “best X for Y” and “X vs Y”. Which means the content formats SEOs have long treated as commodity, like listicles and comparison pages, are, structurally, the formats AI answers are built to consume.
The citation data backs this hard. Seer Interactive's 2026 analysis found listicles to be the single most-cited content type in AI search, around half of top citations. Ahrefs, studying 26,283 URLs cited by ChatGPT across 750 “best X” queries, found 43.8% pointed to “best of” listicles, the pattern holding across Gemini, Perplexity, Claude and AI Overviews.
One caveat that matters, and belongs in any honest version of this argument: the format is not a licence to churn. Google began suppressing self-serving, manufactured listicles in early 2026. What survives is the real thing: transparent methodology, genuine alternatives, comparison tables with verifiable data, external validation. The machine rewards the shape of a good listicle only when there’s substance inside it.
Everything so far treats the fan-out as evidence of a problem. Turn it over and it becomes the most useful planning input you have, because the model isn't only running searches. It's telling you exactly which sub-audiences it's trying to satisfy, and what each one is worried about.
Go back to the carry-on prompt. To personalise its answer, the model also reached for lines like these:
That isn't a keyword list, it's a segmentation study. Every line names a persona or a job-to-be-done and the exact anxiety attached to it: the crushed suit, the toddler, the downpour, the bag that won't close at the gate. The machine has done the audience research and handed you the findings, phrased the way real buyers actually think. Which makes the fan-out usable far beyond a content calendar.
Each sub-query is a brief with the angle built in. “Packing and compression tips to qualify as carry-on” isn't a product page, it's a how-to that earns trust, and the citation, before anyone is ready to buy.
“Carry-on for family travel with toddlers”, “carry-on with a suit-folding compartment”: a page built for one persona converts that persona far better than a catch-all category page, and it maps one-to-one onto a query the model is already running.
Fold the recurring angles back into the category page itself, filters, FAQ blocks, comparison modules, buying-guide sections, so it responds to the sub-queries instead of just listing products. That's the difference between a page that ranks and a page an answer can be built from.
The model has just handed you audience segments and the single message that matters to each. That’s paid targeting and ad-copy angles, “fits a folded suit”, “still closes when it’s full”, plus a backlog of social hooks, all without a separate research sprint. One fan-out can feed SEO, category strategy, paid and social at once.
We are not chasing flimsy machine-generated queries. We are using the fan-out as a research tool: to understand a real persona and their pain points, with the information laid out clearly enough to be lifted straight into an answer. We're not choosing between building for the machine or the human, for research or for purchase. We're building one well-structured, persona-led page that serves all of them at once.
For two decades, search volume was a fair map of demand, because demand expressed itself as a human typing a short keyword into a box. That part is changing.
Instead of getting bogged down with traditional volume numbers, look at the questions behind the question, and discover that some of the most valuable pages are the ones the old compass used to throw away.
AI Overviews made visibility almost free, and the click rarer than ever. How one publisher rebuilt its visits.
Read →
Search is moving from links to answers. Here is how to get cited instead of skipped.
Read →
94% of buyers rank their preferred vendors before ever speaking to a seller. Why the dark funnel matters.
Read →This is what giving a damn looks like in practice. Tell us what's not working.