On November 12, 2026, Google moved Search onto Gemini 4 Hydrogen and called it “the biggest model upgrade to Search yet.” It reached every user worldwide on day one, in every language AI Mode supports. No new ranking guide came with it. Google’s Search Central documentation still says AI features need nothing beyond normal SEO fundamentals. What did change is how the model reads a question, how many searches it runs to answer one, and what kind of answer it builds. Those three mechanics decide which pages get linked now, and each one points to concrete work you can do on your own site this week.
1.What Hydrogen Changed in Search
Gemini 4 Hydrogen is the fast model in Google’s Gemini 4 family. Google’s model announcement describes it as a “daily driver for agentic workflows,” built to execute quickly, while Gemini 4 Carbon and Gemini 4 Argon handle long, deep reasoning. That positioning explains why it runs Search: it has to answer billions of queries at low latency, and Google reports it runs about 2.5x faster than Gemini 3.8 Flash. The benchmark figures in the announcement are Google’s own. No independent test has yet checked how completely Search answers multi-part queries.
Robby Stein, Google’s VP of Product for Search, lists four changes in the launch post. First, intent understanding: the model handles multi-part questions, hard-to-read typos and highly specific requests through deeper instruction following and tool calling. Second, query fan-out: it runs simultaneous searches across the subtopics of a question and links “a broader range of sites,” including sites people may not have found before. Third, visual responses: answers now carry more images, structured layouts, comparison tables and “custom tools built for you.” Fourth, language: responses are more nuanced in every supported AI Mode language, whatever language the query was written in.

For a site owner, the second and third changes matter most. Fan-out decides which pages are retrieved at all, and the visual formats decide what the model can lift from a page once it has it. The first change raises the bar indirectly. A user can now type a long, messy, three-part question, so the answer is assembled from several narrow retrievals instead of one broad one. The sections below take those mechanics in order: write for the sub-questions, make sure nothing blocks eligibility, then give the visual formats real material.
2.Write Pages That Answer the Sub-Questions
Query fan-out means Search no longer runs one lookup per prompt. When someone asks “best budget mirrorless camera for travel video under $800 that works with my old Canon lenses,” Hydrogen splits that into subtopics: budget mirrorless bodies, travel video specs, price ceilings, Canon EF adapter compatibility. It searches each one in parallel and then writes one answer with links. Google says this reaches a wider range of sites than before.
That matters because a page no longer has to rank for the whole head query to be cited. A focused guide to EF-to-mirrorless adapter compatibility can be linked inside the camera answer even if it never ranks for “best budget camera.” This is our reading of how fan-out works, not a Google statement, but it follows from Google’s own description: retrieval happens per subtopic. Pages that answer one sub-question clearly and early are easy to retrieve. Pages that bury the answer in the fourth paragraph of a general overview are not.
Start by mapping the sub-questions behind your most important query. Type the full, realistic question into AI Mode, note which subtopics the answer covers and which sites it links, then compare that list against your page’s headings. Each sub-question your audience actually has should get its own clearly labelled passage that answers it in the first sentence or two, then adds the detail. The outline below shows the camera example restructured this way.
H1 Using Canon EF Lenses on a Budget Mirrorless Body
H2 Which mirrorless mounts take EF lenses with an adapter
H2 Does autofocus work through the adapter (and how fast)
H2 Video-specific limits: stabilisation, focus breathing, crop
H2 Bodies under $800 that work well with adapted EF glass
Every heading in that outline names one question a person could ask on its own, and every section opens with the direct answer. Write the passage so it still makes sense if a reader lands on it with no context from the rest of the page. That means naming the product or concept again instead of relying on “it” from two paragraphs earlier, and stating the condition or number right in the sentence. Self-contained passages are easier for a model to quote accurately, and they read better for people who jump in from a link.
To check the result, run the same multi-part query in AI Mode a few days after the page is recrawled and look at the linked sources for each part of the answer. You can request a recrawl with URL Inspection in Search Console. If a competitor is cited for a subtopic you cover, compare the two passages: theirs usually answers sooner, names the entity explicitly, or includes a number yours leaves out.
3.Keep Your Pages Eligible
Google’s AI features documentation is plain about the requirement: to appear in AI Overviews or AI Mode, a page must be indexed and eligible to show in Search with a snippet. There is no extra AI-specific technical requirement, file or schema. That makes eligibility the easiest place to lose visibility without noticing. One wrong robots rule, CDN setting or preview directive removes a page from every AI answer while normal rankings look fine.

The controls below come from Google’s AI features documentation. Each one changes what Search can show, and a few of them are often misread. Google-Extended is the most common example: it limits training and grounding in some of Google’s other systems, and Google lists it apart from the Search controls.
| Control | What it does to AI features in Search | When to use it |
|---|---|---|
| robots.txt (Googlebot) | Blocks crawling for Search, including AI Overviews and AI Mode | Only for pages you want out of Search entirely |
| noindex | Removes the page from Search and its AI features | Thin, duplicate or private pages |
| nosnippet | No snippet, so not usable as a source in AI answers | Rarely; it removes the page from AI answers |
| data-nosnippet | Hides one element from snippets and AI use | Sensitive blocks such as prices you can’t show |
| max-snippet | Caps snippet length, which limits what can be quoted | When you need a hard quote limit |
| Google-Extended | Limits training and grounding in other Google systems | Separate decision from Search visibility |
Work through the checklist on your highest-value pages. Open robots.txt and confirm Googlebot is not disallowed from the sections that matter. Check your CDN or hosting bot-protection rules, because firewall challenges can block Googlebot even when robots.txt allows it. Search your templates for nosnippet and max-snippet values someone added years ago. Then run URL Inspection on each key URL and use the live test to see the page as Googlebot renders it.
The check is simple: URL Inspection should report the page as indexed, and the live test screenshot and rendered HTML should show your main content as text. If the main answer only appears after a click, a login or a client-side script that fails in the test, the model has nothing to retrieve. Google notes that preview-control changes can take days to months to show after recrawl, so fix eligibility issues first and expect a lag.
4.Give the Visual Answers Something to Use
Hydrogen’s answers now include images, structured responses, comparison tables and small custom tools. A model can only build a comparison table from data that exists somewhere in a clear form. A page that already holds that comparison as a real HTML table, with consistent columns and units, gives the answer something accurate to draw from. A comparison scattered across five paragraphs of prose forces the model to reconstruct it, and it may take another site’s cleaner version instead.
This lines up with Google’s general best practices for AI features: keep important content in text, support it with relevant images and video, and make sure structured data matches what is visible on the page. Nothing here is new. The rollout of more visual formats just raises the cost of skipping it. A comparison saved as a screenshot is invisible as data. Product schema that claims a price the page doesn’t show is a mismatch Google tells you to avoid.
Go through the pages where readers compare options and move each comparison into an HTML table. Use one row per option and one column per attribute, put units in the header, and avoid merged cells. Add original images (your own photos, screenshots or diagrams) with factual alt text that says what the image shows. Keep any structured data you already use in sync with the visible text. Here is the same comparison written both ways.
Before (prose):
Body A is $699 and weighs 450 g, while Body B costs a bit more at $749
but is lighter, and Body C sits at $679 with the best adapter support.
After (table):
| Body | Price (USD) | Weight (g) | EF adapter AF |
| Body A | 699 | 450 | Partial |
| Body B | 749 | 405 | Partial |
| Body C | 679 | 375 | Full |
The bodies and figures in that example are placeholders, not real products. The point is the shape. The table version can be read, compared and quoted row by row without interpretation.
To confirm the work, view the page source and check that the table is real `
5.Measure What Changed
Search Console does not separate AI Mode or AI Overviews traffic from the rest. Google’s documentation says that traffic is included in the Performance report under the “Web” search type, so the November 12 rollout shows up only as a shift inside your normal numbers. That makes the measurement plan more important than any one dashboard. Without a baseline from before the change, you cannot tell a fan-out effect from seasonal noise.
Compare the four weeks after November 12 with the four weeks before it, at the page level. Focus on the pages you restructured around sub-questions. Look for pages gaining impressions on longer, multi-part queries they never ranked for, which is the pattern fan-out should produce. Pair that with Google Analytics engagement for the same landing pages. Google says clicks from pages with AI Overviews tend to come from people who stay longer, so time on page and conversions can tell you more than raw clicks. Keep a short log of the changes you shipped and their dates, so a later jump can be traced to a specific fix rather than guessed at.


