The step your customer never sees
When someone asks ChatGPT, Claude, Gemini, or Perplexity a real buying question, it usually does not answer from memory alone. In the fraction of a second before the reply appears, the assistant quietly runs a handful of ordinary web searches, reads what comes back, and writes its answer from those pages. Your prospect never sees the queries. They see a confident paragraph and, sometimes, a short list of sources. But behind that clean answer is a burst of background searching that looks a lot like the traditional search engine you already know.
This is the mechanic that most brands miss. The assistant is not inventing its recommendations out of thin air. For any question fresh or specific enough to need current information, it is grounding the answer in live search results, the same retrieval-augmented process that powers AI Overviews and every other AI answer that cites the web. One question in can become several searches out. And the pages that show up for those searches are the pages that shape what the assistant tells your buyer.
Why the hidden searches decide who gets cited
Here is the uncomfortable part. If the assistant grounds its answer in web results, then the brands that rank for those background searches are the brands that get read, named, and cited. The ones that do not rank for them are invisible to the model, no matter how good their homepage is. The AI is not choosing the loudest brand or the biggest one. It is choosing from the pages its background searches surfaced.
You are not optimizing for the question your customer typed. You are optimizing for the searches the assistant ran on their behalf, which are frequently not the same words at all.
That reframes the whole problem. You are not optimizing for the question your customer typed. You are optimizing for the searches the assistant ran on their behalf, which are frequently not the same words at all.
Keyword lists model what humans type. Fan-outs model what AI asks.
A traditional keyword list captures how a person phrases things: short, blunt, often ambiguous. But when an assistant decomposes that same intent, it tends to translate one human question into a spread of cleaner, more specific sub-queries. Ask it for the best CRM for a small business and it may quietly search small-business CRM pricing, easiest CRM to set up, CRM versus spreadsheet, and CRM integrations for QuickBooks. Your keyword research would never have surfaced that exact spread, because that is not how your customer talks. It is how the machine reinterprets what your customer wants. Optimize only for the human phrasing and you answer the question the buyer asked while missing every question the assistant actually searched.
Not every question triggers a search
The distinction that matters most, and the one almost nobody makes, is this: not every prompt grounds. When a model already holds a stable, well-settled answer in its trained knowledge, it does not bother searching. Ask it to define a common term or explain a timeless concept and it answers from memory, no live queries, nothing for you to win.
Fresh, specific, or commercial questions are different. Who is the best provider right now, what does something cost this year, which tool integrates with which platform: those push the model to search, because its training data is stale or incomplete. Those are the grounded prompts, and grounded prompts are the only ones that are addressable by anything you publish. If a question never triggers a search, no amount of content changes the answer. If it does, the background searches are your way in. Knowing which of your buyers' questions ground and which do not is the first real map of where GEO can move the needle and where it simply cannot.
This is also why generic advice to just publish more helpful content underperforms. Volume aimed at questions the models answer from memory is wasted; the same effort pointed at grounded, commercial questions compounds. The winning move is not more content, it is content aimed at the exact searches that fire. Getting there means seeing the fan-out for yourself rather than guessing which of your questions belong in which bucket.
The public signal: change one search parameter, watch AI answers move
If you doubt that AI answers ride on ordinary search rankings, late 2025 delivered a natural experiment. Google removed a long-standing search parameter that many tools had relied on to pull one hundred results per page. It was widely reported and publicly documented across the SEO community that, almost overnight, the mix of pages cited in AI answers shifted, and analysts observed unusual movement in the sources these engines leaned on. We are not going to put a number on it, because the credible numbers belong to the analysts who ran them. But the direction of the signal is the point: a change to how traditional search returns results rippled straight into what AI assistants cited. That kind of correlation suggests AI answers draw heavily on overlapping retrieval signals, each assistant maintains its own search stack, but the citation patterns often move together. When the search layer moves, the answer layer tends to move with it.
What to do about it
The practical program follows directly from the mechanic. First, capture your real fan-outs, per persona. Different buyers ask differently, and each phrasing can trigger a different spread of background searches, so model the questions your actual segments ask rather than one generic prompt. Second, gap-analyze: line up the searches the engines really ran against the pages you already have, and find the sub-questions you never answered. Those gaps are your content roadmap, ranked by the queries that keep surfacing.
Third, create and earn the placements that win those searches. Some of that is your own answer-first content. Much of it is not. Because the assistant reads whatever ranks, a strong third-party page that mentions you can be as valuable as your own, sometimes more. Multi-site visibility counts: a top-ranking review, directory, or industry roundup that names you is a legitimate path into the answer, and often an easier one to win than a competitive head term. The goal is not just to rank yourself but to be present wherever the background searches land.
Treat it as an ongoing loop rather than a one-time audit. The searches an assistant runs drift as its models update, as the web changes, and as competitors publish. Re-capture the fan-out on a cadence, watch which sub-questions newly appear or fall away, and keep pointing your content and outreach at whatever is surfacing now. That blend of owned content, structured data, and earned third-party citations is exactly the work our Generative Engine Optimization program runs, and it is the same discipline behind tracking how assistants describe you over time in our AI visibility practice.
See your own hidden searches
You do not have to take the mechanic on faith. Our free GEO Query Fan-Out tool captures the real background searches four engines report when they ground a question: Claude through Anthropic's web-search tool blocks, ChatGPT through the OpenAI Responses API, Gemini through its Google Search grounding metadata, and Perplexity through its search results and citations. It builds buyer personas with the conversational prompts each would actually type, captures the live fan-out across all four engines, enriches it with an AlsoAsked question tree, layers in search volumes, competitor data, and audience signals, then hands you a plain-English action plan. When a prompt answers from memory and runs no search, the report marks that row honestly, so you always know which queries were truly captured and which questions the models never bothered to look up.
It is free and bring-your-own-key: your API keys are used once for your request and never stored. Want to look before you run it? See a sample report, then run your own. Open the GEO Query Fan-Out tool and find out what the assistants are really searching before they answer for you.