Google's AI Mode Fires Off Queries You Never Typed. That's the Part That Matters.

A client sent me a screenshot last month with the subject line "are we in this?" It was Google's AI Mode answering a question about their product category. Three companies got named in the answer. They weren't one of them, despite ranking third for the obvious keyword.

I couldn't tell them why. Not because I didn't look. Because the question they were asking has no answer you can look up.

Here's the reason, and it's the one thing about Google AI Mode SEO that most of the advice out there skips: AI Mode doesn't run the query the user typed. It breaks it into a batch of other queries, runs those in parallel, and writes an answer out of whatever comes back. Google documents this. It's called query fan-out, and once you understand it, most of the "optimize for AI Mode" advice on LinkedIn stops making sense.

I've written before about how AI Overviews eat clicks and what still works when your traffic is down a third. This is a different post. This one is about AI Mode as its own surface: how the retrieval actually works, whether it really is becoming the default, and what your reporting can and cannot tell you about it.

AI Mode Is a Different Surface, Not a Bigger Answer Box

Start with the thing people conflate. An AI Overview is a block on the normal results page. AI Mode is a separate destination, reached through its own tab, the search bar, or Circle to Search, sitting alongside Images and Videos.

That difference isn't cosmetic. Semrush's analysis of AI Mode found that roughly 92% of AI Mode responses show a sidebar with about seven unique domains, and only about 53% of those domains also appear in Google's organic top 10. At the URL level the overlap drops to around 35%. Their query sample there was small, five high-intent queries, so treat the exact figures as directional rather than gospel. But the direction is consistent with what Marie Haynes found doing the same comparison by hand: run one question through classic search and through AI Mode and you often get a materially different set of sites.

Then there's the follow-up. AI Mode keeps context and invites the next question. So you are not competing for a query, you are competing for a position inside a session that may run five turns deep, where turns two through five are shaped by an answer you had no part in.

Query Fan-Out Is the Whole Mechanic

Google's own developer documentation says it plainly: "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." Google's Search blog puts it the same way, describing AI Mode "breaking down your question into subtopics and issuing a multitude of queries simultaneously."

That's the documented part. Past that we're in patent-reading and observation territory, so I'll label it as such. iPullRank's teardown points at Google patents including US12158907B1 ("Thematic Search") and describes eight kinds of synthetic query the system can generate: rephrasings, follow-ups, broader versions, narrower versions, canonical phrasings, translations, logically implied questions, and clarifications. Nobody outside Google knows how many fire on a given search. Estimates in the industry range from a handful to dozens, and anybody quoting you a precise number is guessing.

Here's the practical consequence, and it's uncomfortable.

You are not competing for the query your keyword tool shows you. You are competing to be retrieved for sub-queries the user never typed and you will never see. There is no report anywhere, in Search Console or any third party tool, that lists the fan-out queries that produced an answer you appeared in or missed. Your third position for "best X for Y" might be irrelevant if the fan-out that mattered was "how long does X last" and your page never addresses it.

This is why "rank higher" is an incomplete answer for AI Mode. Ranking is still the entry ticket, Google's retrieval pulls from the same index, but the thing being matched is narrower and stranger than a keyword.

About That "Default" Claim

Every second post about this says AI Mode is now the default Google experience. Let me be precise, because the evidence doesn't quite say that.

What's documented: AI Mode started as a Labs experiment in March 2025, opened to all US users that May, and by October 2025 had expanded to more than 40 countries and 35 languages. At I/O in May 2026 Google rolled its newer model out as the default model in AI Mode globally. On the Q2 2026 earnings call, Pichai said that since expanding AI Mode globally last October, it has "surpassed 1 billion monthly active users."

What Google actually says about default status is narrower. Covering I/O 2026, TechCrunch noted that "while Google says that AI Mode is not the default experience, Search's user interface encourages users to ask follow-up questions instead of scrolling down to the links." That's the honest version: not the default, heavily nudged toward.

And then there's a discrepancy nobody has reconciled. Datos clickstream data for Q1 2026 put AI Mode at under 0.2% of desktop search events in the US. One billion monthly actives and a fraction of a percent of desktop sessions can both be true if AI Mode usage is overwhelmingly mobile, shallow, and driven by placement rather than habit. But I can't prove that, and neither can the people confidently telling you AI Mode is where search happens now.

So: growing fast, prominently placed, technically not the default, and a much smaller share of actual desktop search behaviour than the headline user number suggests. If a consultant quotes you the billion and skips the rest, that's a tell.

What Search Console Actually Tells You About AI Mode

This is the part that matters if you report to anyone, and it's where I've had the most awkward client conversations.

AI Mode data started counting toward Search Console totals on June 17, 2025, folded into the Web search type with no separate dimension. It was in your numbers for a year before most people knew it was there. Not hidden, exactly, just indistinguishable.

On June 3, 2026, Google launched dedicated generative AI performance reports. Real progress, and I went through what my own AI numbers looked like when that reporting landed. But read what they give you: impressions, broken down by page, country, device and date. They do not break out clicks or click-through rate by surface. Brodie Clark's testing found the same thing, impressions for pages appearing in AI Mode, no clicks and no query data at that level.

Put that in client language. You can say "we appeared in AI features this often." You cannot say "AI Mode sent us this many visits," and you cannot say which questions triggered it. Anyone selling you an AI Mode traffic number built from Search Console is doing arithmetic on data that doesn't exist. Layer on the fact that Google now routes result links through a redirect and the referrer picture gets murkier still.

I've changed how I report because of this. AI feature impressions go in as a visibility indicator with an explicit caveat that it is not traffic. Then leads and revenue carry the actual argument. That's less satisfying than a clean chart. It's also the only version that's true.

What This Actually Changes in the Work

Now the deflating part, and I'd rather deflate you than sell you a new discipline.

The whole of Google AI Mode SEO, according to Google, is that there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." I went through their full AI search guidance and what it means for the GEO crowd and the short version holds: it's SEO. No llms.txt, no chunking rituals, no AI-specific rewrite.

What fan-out genuinely changes is emphasis, in two places.

Passage-level retrievability. The unit being matched is a passage answering a specific sub-question, not a page matching a keyword. A page that buries the direct answer under 600 words of preamble is harder to retrieve from than one that answers plainly under a clear heading. Same page, same ranking, different odds.

Explicit coverage of adjacent questions. If the fan-out includes "how much does it cost," "how long does it take," and "what goes wrong," a page that only sells the service can't be pulled for any of them. Answering those questions on the page, in their own words, is the whole tactic. It is also just good content, which is my point.

That's it. That's the list. Everything else is the same job: be indexed, be genuinely useful, be specific, be quotable in one sentence.

What I'm Not Going to Pretend to Know

I don't have AI Mode click data from my client sites, because it doesn't exist to have. Nobody does. Anybody showing you an AI Mode traffic dashboard built it out of proxies.

I don't know whether AI Mode becomes the actual default. Google says it isn't. The interface is clearly being built as though it will be. Those two things have been true simultaneously for a year now.

I don't know whether the sidebar of seven domains ends up sending meaningful traffic or turns into an attribution line. Pichai says Google is sending "billions of clicks to websites every week through AI features." That number has no denominator, so I can't do anything with it.

What I'd act on, and what I'd call the honest core of Google AI Mode SEO: fan-out is real and documented, ranking is still the entry ticket, passage clarity is the cheapest thing you can fix, and your reporting should say "impressions" where it means impressions. That's an unglamorous list. It's what the evidence supports.

If you're staring at an AI Mode answer that names three competitors and not you, I can't promise I'll get you in there. What I can do is tell you honestly which sub-questions your page fails to answer, and whether the rest of your search visibility is healthy underneath it. Send me the site and I'll take a look.