Gemini 3 Churned 42% of AI Overview Citations. Gemini 4 Hydrogen Just Landed Mid-Core Update.

Google replaced the model behind its AI answers on November 12, two days into a core update that is already reshuffling the rankings those answers draw from. The last time Google made this swap, roughly four in ten cited domains lost their slot. This time, nobody can say which change is doing what, and the measurement tools broke during both transitions. Here is what the precedent predicts, what Google’s “more links” promise is actually testable for, and how to separate the two effects before you report to a client.

Published November 13, 2026 · On-Page.ai editorial

Between 10:06 a.m. Pacific on November 10 and sometime on November 12, Google changed three things that all move AI citation data at once. It launched the November 2026 core update, a roughly two-week re-ranking of the whole organic index. It confirmed a data-logging bug in Search Console’s Generative AI performance report covering November 9 to 11. And it made Gemini 4 Hydrogen the default model in AI Mode worldwide and the generator of the hardest questions in AI Overviews.

If you track citations, AI referral traffic, or client rankings, your data for the next two weeks has three possible causes and one corrupted baseline. The agencies that will answer “why did our citations move?” convincingly are the ones that log the dates separately, starting now.

What Google changed, and what it did not say

On X, Google’s Search leadership called Hydrogen “the biggest model improvement to Search yet” (Nick Fox) and said it is “the new default model in AI Mode & being used for your hardest questions in AI Overviews” (Robby Stein, VP of Product). It replaces Gemini 3.8 Flash, completing tasks about 2.5 times as fast, per VentureBeat and Ars Technica.

Google’s announcement makes one claim that matters for websites, in two phrasings. Query fan-out, the technique where one question becomes several simultaneous sub-queries, is “supercharged,” which Google says will connect searchers “to a broader range of helpful links than before” and help them “find sites you previously might not have discovered.” Rajan Patel added that responses come “complete with even more links so you can discover sites from across the web.”

Three things are conspicuously absent, as Search Engine Journal’s Matt G. Southern noted within hours:

• No counts. Neither the Search post nor Stein’s thread says how many more links, or to how many more sites.
• No definition of “hardest questions.” Hydrogen handles the hardest AI Overview questions. Google has not said which questions qualify or what model answers the rest, so nobody knows what share of AI Overviews actually just changed.
• No word on what happens to organic retrieval. Barry Schwartz observed that the swap “only impacts AI Overviews and AI Mode but not core rankings,” which is true but incomplete: the core update running alongside it does move the rankings those AI answers retrieve from. More on that below.

The last brain-swap: what January’s numbers predict

When Google made Gemini 3 the default for AI Overviews on January 27, 2026, SE Ranking tracked 100,000 keywords across 20 niches and measured the citation pool before and after. This is the dataset to hold Hydrogen against.

The churn underneath those totals is the number to remember. 42.4% of previously cited domains vanished from AI Overviews (37,870 domains out), and 51.7% of the post-switch citation pool was made of domains that had never been cited before (46,180 in). But the churn was almost entirely a long-tail event: among the 500 most-cited domains, exactly one disappeared, and none entered. YouTube kept its 10.74% share of all citations, Reddit 4.01%, Facebook 1.85%. Citation concentration, measured with a Herfindahl-Hirschman index, rose 44%. Gemini 3 broadened the base and doubled down on the leaders at the same time.

If Hydrogen repeats the Gemini 3 pattern, the forecast for late November is specific:

• Sites with few AI citations should expect churn in either direction. Sites that are cited heavily should expect stability.
• Per-citation value keeps shrinking as sources per answer grow. Seer Interactive found cited brands still earn 35% more organic clicks and 91% more paid clicks than non-cited brands on the same queries, so citations still pay. They just pay less each.
• Any new “broader range of links” will show up first as more unique domains and a lower overlap between cited pages and the organic top 10, which already sits at 38% per Ahrefs and 19% per SE Ranking depending on method.

Why this week is a harder problem than January

January 27 fell in a quiet window. The December 2025 core update had finished rolling out on December 29, and the next core update was months away. When SE Ranking measured churn, the organic index underneath the AI answers was standing still, so a changed citation pool meant a changed model.

That inference is unavailable this week. AI Overviews ground their answers in the search index: the fan-out sub-queries run against organic rankings, and the results feed the generator. A core update re-ranks that index, so citations move whenever rankings move, even with the identical model. Between November 10 and roughly November 24, any churn you observe in AI Overviews has at least two plausible causes, and your Search Console baseline for the same period was corrupted by the logging bug for its first three days.

There is one built-in instrument in Google’s own announcement, and we noticed it while parsing Stein’s sentence. Hydrogen powers AI Mode entirely but only the hardest AI Overview questions. Easier AI Overview queries appear to stay on the previous stack. That gives you a crude control group:

• If citation churn shows up on simple, factual queries and complex ones alike, the retrieval pool is the likely driver, which points at the core update.
• If churn concentrates in complex, multi-part queries (and in AI Mode, which switched wholesale), the model change is the likely driver.

Google has not published how it splits easy from hard, so treat this as a directional test, not a clean experiment. It is still better than reading the month as one blended change.

Day-one field notes: what Hydrogen-era overviews looked like on November 13

We ran a small set of live probes from US locations on the morning of November 13, roughly 18 hours after the switch. Three queries is not a study, and we have no pre-November 12 baseline of our own to compare against, so treat these as observations about what a tracker has to deal with, not as measurements of churn.

The attribution playbook for the next three weeks

This is the sequence that keeps client reporting defensible while the two changes overlap.

The supporting habits, whatever the flowchart says:

• Freeze a query set today. Pick the queries that trigger AI Overviews for your client’s topics, and capture the cited domains (panel sources and inline chips counted separately, “+N” chips expanded where possible) on fixed dates: one capture now, one at rollout completion, one a week after. A single snapshot proves nothing, because AI Overviews vary by session.
• Record the two dates in every client report for November: core update start November 10, model swap November 12, completion date to be logged when the Search Status Dashboard posts it. SEJ’s Southern makes the same recommendation: logging both lets you split the month instead of reading it as one change.
• Do not treat the first churn number anyone publishes this week as real, including ours if we printed one. The Gemini 3 lesson is that the first three weeks of data were contaminated by a concurrent bug, and SE Ranking’s own early conclusion (unique domains down 14%) flipped to the opposite (+9.3%) once the bug cleared. The earliest defensible Hydrogen measurement window opens about a week after the core update completes, in early December.

What would confirm Google’s promise, and what would break it

“A broader range of helpful links” translates into three measurable predictions. When a valid post-rollout dataset exists, these are the tests:

One more watch item: the sourceless rate. Gemini 3’s rollout came with a display bug that left one in ten overviews without sources and settled at 1.27%, eleven times the old baseline. Our one failed generation on day one is a single data point, but if the rate climbs again during this rollout, early churn measurements will be contaminated exactly as they were in January, and Google’s measurement record in these transitions is now two for two. Trackers should publish failure rates alongside churn rates, or their churn numbers are not comparable week to week.

The bottom line

Google shipped its biggest Search model upgrade into the middle of a core update, and both changes move the same citation data your reports depend on. The January precedent says the churn will be real, large in the long tail, and invisible at the top. The honest numbers will not exist until the rollout finishes and a clean week of data accumulates, which puts the first trustworthy Hydrogen citation study in early December. Until then, log the dates, split easy queries from hard ones, exclude November 9 through 11 from any Search Console baseline, and tell clients there are two causes in flight. The agency that can show its work on those four points will be the one clients trust with the answer when the real numbers land.

Sources

• Google Search Gets Major Upgrade With Gemini 4 Hydrogen, Barry Schwartz, Search Engine Roundtable, Nov 12, 2026 (includes Google’s fan-out claims and quotes from Nick Fox, Robby Stein, Rajan Patel).
• Google Makes Gemini 4 Hydrogen The Default Model In AI Mode, Matt G. Southern, Search Engine Journal, Nov 12, 2026 (missing link counts, “hardest questions” gap, prior default changes).
• Gemini 3 Replaces 42% of Previously Cited Domains in AI Overviews, SE Ranking, Feb 26, 2026 (100,000-keyword study, bug timeline, methodology).
• Google November 2026 core update rolling out now, Search Engine Land, Nov 10, 2026; Search Engine Roundtable coverage; Search Engine Journal (dashboard start 10:06 a.m. PT).
• Google confirms Generative AI performance report logging bug, Nov 9–11, Search Engine Roundtable.
• Google rolls out Gemini 4 Hydrogen, replacing 3.8 Flash, Ars Technica, Nov 12, 2026; VentureBeat on speed and cost.
• Update: 38% of AI Overview Citations Pull From The Top 10, Ahrefs, Mar 2026 (863K SERPs); AI Overviews Study: 25+ Stats From 405,576 Google Searches, Surfer, Nov 5, 2026.
• Gemini 3 Cut AI Overview Citations: The 4-Step Recovery Playbook, Frase (Seer Interactive click data, top-10 overlap figures).
• Google Search Status Dashboard, ranking history (December 2025 core update completed Dec 29; November 2026 start).
• Live SERP probes captured by On-Page.ai on Nov 13, 2026, ~10:25 a.m. ET from US locations. Full-page screenshots retained in our research archive.