Because Google no longer answers your question with one search. In AI Mode, Google splits your question into smaller sub-questions and runs them all at once. That process is called query fan-out. Your page is not competing for one query anymore. It is competing for the many hidden ones.
This is the single biggest shift we deal with when clients ask us why their traffic looks strange in 2026. Rankings look fine. Impressions look fine. Clicks do not. The reason is that the search that brought the visitor was never the search you optimized for.
We build Webflow sites for founders and marketers who need to be found. So we spend a lot of time reading what Google actually publishes about how its systems work, rather than what the SEO rumor mill says. Query fan-out is one of the few mechanics Google has described in plain language, and it is worth understanding properly.
Query fan-out is Google's method of breaking one question into subtopics and running many searches at the same time. Google describes it on its own blog as a "query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf." The AI then reads those results and writes one answer.
Google introduced the term in a Search blog post published on 20 May 2025 covering the AI Mode updates from Google I/O. That post is still the clearest official description of the mechanic anywhere.
The important word is "simultaneously". These are not follow-up searches that happen after you click something. They happen in parallel, in the moment, before you ever see an answer. A single question can pull sources from many different pages, and each of those pages was retrieved for a different reason.
Google applies the same idea to images. In the same body of work it describes a visual fan-out that breaks one image search into many sub-queries to understand what is actually in the picture. And in Deep Search, Google says the technique goes further still, issuing hundreds of searches to build a fully cited report.
Because people started asking longer questions. Google reported in May 2026 that a typical AI Mode search is triple the length of a traditional Search query. A long, messy question does not match one page well. Splitting it into parts and searching each part separately produces a better answer than trying to match the whole thing.
The scale behind this is real. In that same May 2026 update, Google said AI Mode had passed a billion monthly active users and that AI Mode queries had more than doubled every quarter since launch. Google also said more than one in six searches in the United States now use voice or images.
Those are Google's own numbers about Google's own product, which is exactly the kind of source worth building strategy on. We would not make a plan around a third-party estimate of AI Mode usage. We will make a plan around what the company running the system says it is seeing.
It moves the unit of work from the keyword to the question cluster. Instead of picking one phrase and writing one page for it, you map every sub-question a reasonable person would need answered on the way to a decision. Then you make sure your page answers each of those sub-questions clearly and separately.
In practice we now build a page outline by writing out the questions first. If the topic is "best CRM for a small agency", the fan-out around it almost certainly includes pricing, setup time, integrations, team size limits, and migration effort. A page that only sells one product answers one of those. A page that answers all five can be pulled into the response many times over.
This is a change in how you spend research time, not how much you spend. Our approach to keyword research now starts with the sub-questions and works backward to the head term, rather than the other way around.
A page that is indexed, readable as text, and answers a specific sub-question in a self-contained way. Google's Search Central documentation is direct about the requirement: to be eligible as a supporting link in AI Overviews or AI Mode, "a page must be indexed and eligible to be shown in Google Search with a snippet."
That sentence rules out a lot of sites without anyone realizing it. If your content only appears after JavaScript runs, or your page blocks snippets, you are not in the pool at all. We see this constantly on sites built as client-side apps, and it is why we prerender our own blog to static HTML rather than fetching posts in the browser. It is also why we wrote about why AI crawlers miss JavaScript-rendered content.
The second half matters just as much. A sub-question retrieval wants a passage that stands on its own. If the answer to "how long does migration take" is spread across four paragraphs and depends on context from earlier in the page, it is a weaker candidate than a tight paragraph that answers it outright.
No. Google says so plainly. Its AI features documentation states that "you don't need to create new machine readable files, AI text files, or markup to appear in these features" and that there is "no special schema.org structured data that you need to add." It also says there are "no additional technical requirements" beyond normal indexing.
We think this is the most ignored sentence in AI search right now. A whole industry of AI-visibility products implies there is a secret file or a hidden tag. Google's own developer documentation says the opposite. Standard structured data still helps, and Google asks that structured data match the visible text on the page, but it is not a backdoor into AI Mode.
Our position is that the work is boring and it is the same work as always. Get the content indexable. Get it into the HTML. Make the important content available in textual form, which is language Google uses directly. Then make the answers good.
Give every sub-question its own heading, and answer it in the first forty to sixty words underneath. Depth comes after the answer, not before it. This is the same structure that wins featured snippets, and it happens to be the structure a fan-out retrieval can slice cleanly.
We write our headings as questions on purpose. A heading that reads "Pricing" tells a retrieval system very little. A heading that reads "How much does a Webflow migration cost?" matches the shape of a sub-query, and the paragraph under it can be lifted as a standalone answer. That is the whole game.
Consistency in wording also matters more than it used to. If you call it "Core Web Vitals" in one section and "speed scores" in another, you have split your own topic into two weaker signals. Pick the term the documentation uses and use it every time.
The related habit is covering a topic completely rather than in fragments. Fan-out rewards sites that can answer the whole cluster, which is really just another argument for building topical authority instead of chasing single keywords.
Better pages, almost always. Fan-out does not reward volume. It rewards a page that can satisfy several sub-questions at once, because that page gets retrieved several times for the same user question. Ten thin pages each answering one sub-question compete with each other and win less often than one thorough page.
We take a firm position here because we have watched the alternative fail. Publishing a page per long-tail phrase was a workable tactic when retrieval matched one query to one document. When one question spawns many retrievals, the thorough page is simply worth more per unit of effort.
There is a limit. A page trying to answer thirty unrelated sub-questions becomes a page about nothing. The judgment call is whether the sub-questions belong to the same decision. If a reader would ask all of them in one sitting, they belong together.
Mostly through Google Search Console, with realistic expectations. Google states that traffic from AI features appears in the Search Console Performance report under the "Web" search type, mixed in with the rest of search data rather than broken out separately. There is no AI Mode filter.
So you cannot get a clean number. What you can do is watch for the pattern: impressions holding or rising while clicks fall on informational pages, which is the signature of an answer being read without a visit. That is not a failure by itself. Being the cited source in an answer still builds recognition, and it still sends the visitors who need more than a summary.
Outside Google, the honest method is manual. Ask ChatGPT, Perplexity, and Claude the questions your customers ask, and see whether you appear. It does not scale, but it is real data rather than a vendor's estimate. We pair it with the tracking approach in our guide to optimizing for Google AI Overviews.
For the reader, using a structure the fan-out can read. These are not in conflict. A person scanning a page wants a clear question and a direct answer under it. A retrieval system wants the same thing. The structure that helps one helps the other.
Where they do conflict, the reader wins. We have seen pages so heavily shaped for machines that they read like a form. Those pages get skimmed and abandoned, and a page nobody finishes is a page nobody links to or trusts. Google's own guidance keeps pointing back at making content genuinely useful, which is not a slogan so much as an admission that they have no better proxy.
Pick your three most valuable pages and rewrite the headings as real questions, with a direct answer in the first two sentences under each. Then confirm the page renders its content in HTML without JavaScript. Those two changes do more than any AI visibility tool you can buy.
After that, look at coverage. Take one topic you care about and write down every sub-question a buyer would ask. If your site answers four of twelve, that gap is your content plan. It is a more useful plan than a keyword list because it maps to how the answer actually gets assembled.
If you want a hand working out which pages are worth the effort, or you are staring at a site that renders everything in JavaScript and wondering how much trouble you are in, we are happy to walk through it. Let's talk. You can reach our team at phoenix.studio.
Tell us where you want to go. We'll tell you how we'd get you there.