Why AI Mode May Never Cite the Same Page AI Overviews Do

Cited in one, invisible in the other

Say you did the work. You read through what actually blocks a page from AI Overviews, checked your response headers, cleaned up a stray directive, and confirmed it: type your core Medicare or ACA question into Google and there you are, named in the AI Overview panel above the ten blue links. You'd be forgiven for thinking the AI citation problem is solved.

Then a client mentions they asked Google something on their phone and landed in a full-screen chat thread instead of a search results page. Or you open the Google app yourself, tap into the tab labeled AI Mode, and ask the exact same question you're cited for. A different answer comes back. Different sources. You're not in it.

That's not a bug and it's not a personalization quirk. It's two different Google products, built by two different teams inside the same search organization, pulling from two citation pools that barely touch. Google says as much in its own documentation, in one sentence easy to skim past: AI Mode and AI Overviews "may use different models and techniques, so the set of responses and links they show will vary"¹. Most agencies read past that line because it sounds like a minor technical footnote. The research on how much they actually vary says it's closer to the whole story.

If you've already worked through what changed in AI citation this year or you're relying on rankings to tell you whether your AI search traffic even exists, this is the layer most of that work never touches: a second, separate surface with its own separate list of who gets named, growing faster than the one everybody's been optimizing for.

The honest version of the problem is that most of the AEO advice circulating right now, including plenty of what we've published, was written back when "AI citation" mostly meant one thing: the summary panel sitting above search results. That was a reasonable place to start, since AI Overviews got there first and was the surface everyone could actually see and screenshot. AI Mode spent its first year as a smaller, opt-in feature most agencies never noticed, which is exactly why the research on how differently it behaves is only now catching up to how big it's actually gotten.

AI Mode is not AI Overviews with a new name

AI Mode launched in the United States in May 2025 as a limited, opt-in experiment. By May 2026, Google's own report on the feature said it had surpassed a billion monthly active users globally, with query volume more than doubling every quarter since launch³. That's not a beta feature anymore. It's a full search product with its own dedicated tab in Google Search and the Google app, and it works nothing like the panel sitting above your regular results.

AI Overviews sits inline, above the ten blue links, inside a search you were always going to run anyway. You can ignore it and scroll straight to the organic results underneath. AI Mode replaces that experience entirely: there's no ranked list under the answer, just a conversational thread you can keep questioning, the way you'd keep talking to an assistant rather than re-typing a search box ten times³.

The query behavior backs that up. Google's own data shows the average AI Mode search runs three times the length of a typical Google query, more than one in six U.S. searches inside AI Mode now include voice or an image, and planning-style questions, the kind that start with "where should I" or "ideas for," grew 80% faster over six months than AI Mode's overall query volume³. Read that against what a Medicare or ACA shopper actually asks: not a three-word keyword, but something closer to "I'm turning 65 in March and my current job's insurance ends, what do I need to figure out first." That's exactly the query shape AI Mode is built for, and exactly the query shape a ten-word landing page never answers in one place.

Notice the row that's identical on both sides: retrieval technique. Both surfaces run on query fan-out, the same mechanism, described the same way in the same Google documentation¹. That similarity is exactly why it's tempting to assume the two surfaces behave the same way and cite the same pages. The data says the shared mechanism produces two almost entirely different outcomes.

The voice and image share of AI Mode queries matters here too, since it's one more way the input itself differs from a typed search box. More than one in six U.S. searches inside AI Mode now include voice or an image³, which means a real, if smaller, share of your prospects are photographing a plan document or a bill and asking what it means, out loud, in the car or at the kitchen table, rather than typing a phrase into a box at all. None of that changes what wins the citation. It does mean the "query" your page has to answer was never a keyword to begin with. It was always a sentence someone said the way they'd say it to a person.

The 13.7 percent number

Ahrefs published the most direct comparison of the two surfaces to date on December 15, 2025. Analysts Despina Gavoyannis and Xibeijia Guan ran roughly 540,000 query pairs through both AI Overviews and AI Mode to compare which URLs each one cited, and a further 730,000 query pairs to compare how similar the actual written answers were, using data pulled in September 2025⁴. The headline finding: only 13.7% of the URLs cited overlapped between the two surfaces for the same query. Narrow that to just the top three citations each surface showed, and the overlap rose only slightly, to 16.3%⁴.

Here's the part that makes the number strange instead of just low. The two surfaces largely agreed on the substance of the answer. Average semantic similarity between AI Overview and AI Mode responses to the same query came in at 86%, and 89.7% of the response pairs scored above 0.8 on a cosine similarity scale, a standard measure of how closely two pieces of text match in meaning⁴. In plain terms: ask the same question in both places and you'll usually get close to the same conclusion, delivered by an almost entirely different set of messengers.

Translate that into an agency's own book of business. If your top five Medicare or ACA landing pages are all cited, correctly and consistently, inside AI Overviews, Ahrefs' data says the honest expectation for AI Mode is that roughly four of those five citations don't carry over at all. Not because the pages are wrong. Because AI Mode is, functionally, a different publication asking a different set of sub-questions and keeping its own separate list of who answered them well.

Why fan-out decides who gets named

Both surfaces run on the same underlying mechanic, and Google defines it in one sentence: query fan-out is "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query"². Instead of matching your page to the one phrase a prospect typed, the model spins that phrase into several related sub-questions behind the scenes and goes looking for a source for each piece separately. Your page doesn't need to win one query. It needs to win several, asked by a system that never shows you which ones it asked.

Surfer SEO put a number on how much that breadth matters, in a study published December 18, 2025. Analysts pulled 10,000 keywords, extracted 173,902 ranking URLs and 33,000 distinct fan-out sub-queries generated for those keywords, and found that 76% of the original 10,000 keywords triggered an AI Overview at all⁵. The core finding: pages ranking for multiple fan-out sub-queries, not just the original keyword, were 161% more likely to get cited than pages ranking for the main query alone, a relationship the researchers measured at a Spearman correlation of 0.77, a strong statistical link⁵.

Put those two studies together and the overlap number stops looking mysterious. AI Overviews and AI Mode both fan a question out into sub-queries, but nothing says the two systems generate the same sub-queries, in the same order, weighted the same way. AI Mode's average question already runs three times longer and often continues across follow-up turns, each one triggering its own fresh fan-out³. A page can win the exact sub-query slice AI Overviews happened to run, and simply never surface for the different slice AI Mode ran for what looks, on the surface, like the identical question.

The practical read for an agency site isn't complicated, even if the mechanism underneath it is. A single page that thoroughly answers a topic and its adjacent questions has more fan-out slices it can win, in either system, than ten narrow pages each built around one keyword variant. Google's own guidance says the narrow-page approach doesn't just underperform, it's a spam-policy risk, not a growth tactic².

Worth defining the three terms this whole guide sits on top of, since they get used loosely enough that it's easy to lose track of which one you're actually working on. SEO is being found in classic search results at all, the ranking, indexing, and speed work underneath everything else. AEO, answer engine optimization, is being the passage an AI system lifts and attributes when it builds an answer. GEO, generative engine optimization, is writing that passage so it stays accurate, sourced, and quotable once it's extracted and rephrased somewhere else entirely. Fan-out coverage sits inside AEO and GEO both: it's the property that decides whether your passage gets pulled into the answer in the first place, on either surface.

What this looks like by product line

Fan-out and low citation overlap are abstract until you run them against the actual questions your prospects ask, which look different depending on what you sell. Three examples, one per product line, show why a single narrow landing page keeps losing to a page built around the whole question.

Medicare. A prospect turning 65 rarely asks one question. Inside an AI Mode thread it might run: "what do I need to do before I turn 65 in March," followed by "does that change if I'm still working," then "what happens if I miss the window." Google's own fan-out definition means each of those is treated as its own retrieval pass, pulling from whichever eligible pages answer that specific piece². A page that covers only the Initial Enrollment Period, with nothing on working past 65 or the late-enrollment penalty, wins the first fan-out slice and loses the next two, in the same conversation, to whichever competitor's page does cover them.

ACA and health. The planning-question growth Google measured, "where should I," "ideas for," questions that grew 80% faster than AI Mode's overall query volume, shows up here almost word for word³. "Where should I look for a plan that covers my current doctor" is a planning question, not a keyword. It fans out into coverage area, network participation, and subsidy eligibility as three separate sub-questions, and a subsidy calculator page that never mentions provider networks only ever wins one of the three.

Life and final expense. These conversations skew even further into the multi-turn format AI Mode was built for: "how much life insurance do I actually need," followed by "what's the difference between term and whole life for someone my age," followed by "how does underwriting work if I have a health condition." Three follow-ups, three fresh fan-out passes, three separate chances to be named or skipped, all inside one conversation a single narrow page can't keep up with turn by turn.

What it costs while you're not looking

The backdrop to all of this is a search landscape that increasingly never sends a click at all. SparkToro's Rand Fishkin analyzed Similarweb's U.S. desktop and mobile clickstream panel for January through April 2026 and found 68.01% of Google searches ended without a click, published June 9, 2026⁶. That's not a click going to a competitor instead of you. It's a search that used to become a website visit for someone, and now more often becomes an answer, read and acted on, with no visit at all.

Insurance buyers are already inside that shift. J.D. Power's AI Insurance Experience Study, fielded June through July 2026, found nearly one in three auto and home insurance customers, 29%, already use AI tools to research coverage, service a policy, or shop for a new one⁷. Among customers who used AI specifically to research products or coverage, 37% changed their policy based on what the AI told them. Among those who used AI to shop for a new policy, 42% went on to purchase one⁷. Those aren't people idly browsing. They're people whose next carrier or agent gets shaped, in part, by whichever source an AI system happened to name, before any human conversation ever starts.

Now run the two studies together. A meaningful share of your prospects already form a decision from an AI answer before they call anyone. That AI answer increasingly comes from a system, AI Mode, with over a billion monthly users and query volume still doubling roughly every quarter³. And whatever citation work already earned you a mention in the older AI Overview panel carries over to that faster-growing surface only about 14% of the time⁴. Optimizing for one surface and assuming it covers the other isn't a small gap. It's closer to running half the race.

Put a rough number on it using your own analytics. Pull your current monthly organic sessions on the one page that answers your single most-asked question, the one a prospect types or asks during AEP or open enrollment. Whatever that number is today, the honest read from the research above is that a shrinking share of it arrives as a click at all, SparkToro's 68.01% zero-click figure says as much across the board⁶, and of the share that does turn into an AI answer instead, being cited in AI Overviews buys you a citation shot in AI Mode roughly 14% of the time, not automatically. That's not a reason to stop measuring rankings. It's a reason to stop treating rankings as the whole measurement.

How to check your own site in AI Mode

None of this requires a paid monitoring tool to start finding out where you stand. It requires about twenty minutes and a spreadsheet.

1. Open AI Mode directly, in a private or logged-out window. Results inside AI Mode are generated live per session rather than pulled from a cached, stable list, so a private window keeps your own search history and account from skewing what you see.
2. Ask your real client questions, in their words. Not keyword phrases. Pull five to ten questions straight from actual calls or emails, the way a prospect would actually type or say them. A Medicare agency might start with "what do I need to figure out before I turn 65." An ACA agency might start with "where should I look for a plan that covers my current doctor." A life agency might start with "how much coverage do I actually need for my family." Whatever your book of business actually fields calls about, start there.
3. Log every domain named, not just whether you appear. A simple sheet: date, question asked, every source cited, whether you were one of them. This tells you who's winning the fan-out slices you're not, which is the more useful signal over time.
4. Ask a follow-up question in the same thread. AI Mode's conversational format is exactly what a single AI Overview panel can't test, and each follow-up triggers its own fresh round of fan-out, sometimes naming an entirely different set of sources than the opening answer did.
5. Repeat monthly. Fan-out coverage shifts as competing content changes and as Google's models change, neither of which you control. A one-time check tells you where you stood on one day. A repeated check tells you whether you're gaining or losing ground.

How we build for both surfaces at once

Nobody, including Google, can promise a citation in either surface for a specific query. What the research above does make clear is what widens your odds in both at once, and it isn't two separate strategies. It's one: pages built around the full breadth of what a prospect actually needs to know, sourced inline, instead of thin pages chasing one keyword variant apiece.

That's the build standard behind every site we ship. We build on Astro, which ships zero JavaScript by default, so pages arrive as finished HTML instead of a runtime that has to load, boot, and then draw the content, the way a page-builder platform works. Our builds score 90+ on Lighthouse performance on mobile and desktop, and some score 100. Every site ships a schema stack, an llms.txt, and a sitemap that update themselves the moment a new page publishes, so nothing about the technical side of citation eligibility is left to a plugin default nobody chose on purpose⁸.

The content cadence matters just as much as the build. Digital Foundation's Pro and Scale tiers add one or more new, sourced blog posts every week, each one built to cover a topic's full breadth rather than a single narrow phrase, which is exactly the property both the Ahrefs and Surfer SEO research point to as what actually widens fan-out coverage⁴⁵. Starter, verified live on the pricing page at 247 dollars a month, includes the complete compliant website with AI citation optimization built in from the start, alongside Google Business Profile management and an AI chat widget, on a 14-day free trial⁸.

One more piece of that cadence is worth naming, because it's the difference between a page that widens fan-out coverage and a page that just adds word count. Every figure in a post we publish comes from a primary source pulled and cited the same session the page is written, not recalled from memory or lifted from another site's roundup. That's the same standard this guide was held to: every statistic above traces to a named study or a Google document fetched while writing it. A page built on invented or stale numbers can still rank. It has a much harder time getting quoted by a system that's checking whether a claim actually holds up before repeating it.

We don't run an AI Mode citation guarantee, because nobody honestly can. What we can tell you is that a site built around thorough, sourced answers has a real shot at both surfaces, and a site built around ten thin keyword pages has a real shot at neither, and that difference shows up in the research above regardless of which company builds the site.

What actually changes

Not a guaranteed spot in either AI answer. Nobody can promise that, and anyone who does is guessing. What changes is that you stop treating an AI Overview citation as proof the whole AI citation question is settled, because the research says it settles roughly 14% of it.

You start testing the surface that's actually growing, not just the one that's easiest to check because it sits above results you were already looking at. And the next time a client mentions getting a different answer from Google's AI on their phone, you'll already know why, instead of hearing about a gap in your citation strategy for the first time from someone you're trying to keep as a client.

Questions agents ask