Gemini 4 Hydrogen hits Google Search: what changes

On November 12, 2026, Google launched Gemini 4 Hydrogen, the cheapest model in the new Gemini 4 family. The news came out as a developer topic, but three sentences in the announcement matter to any business owner: Hydrogen becomes the model behind Google Search's AI answers, becomes the free model in the Gemini app, and arrives in Gmail, Drive and Sheets. Below, what changes for those who have a website, for those who use AI at work, and for those who pay for AI inside their own system.

What Google launched

The Gemini 4 family has three models. Carbon and Argon are the strongest, built for long reasoning and hard problems. Hydrogen, announced now, is the fast-execution one: writing, summarizing, organizing, calling tools and repeating tasks at high volume.

On the market's most widely used comparison, the Artificial Analysis Intelligence Index, Hydrogen scored 51 in the default configuration, versus 50 for Argon, according to Google itself. Artificial Analysis placed it just below Claude Sonnet 5.5 and GPT-6.1 Sol. Google also says that, compared with Gemini 3.8 Flash, the previous model in the same tier, Hydrogen finishes tasks 2.5 times faster and makes up claims more than five times less often. These are the company's numbers, measured in its own tests, and it's worth treating them as a promise until independent measurements appear.

What matters for a small business is where it shows up:

• In Google Search, for all users, in all countries.
• In the Gemini app, including the free plan.
• In Gmail, Drive and Google Sheets, through Gemini, with different rules per plan.
• For developers, through the API, with its own pricing.

What changes in Google Search

When someone searches on Google today, some searches show an AI-written answer at the top (the “AI Overview”), and there is also “AI Mode,” a conversation with search. Hydrogen enters in these answers. Google lists three changes:

1. Better understanding of questions as people actually speak. Multi-part questions, typos or very specific requests become understood. In practice, searches like “flower shop that delivers on Sunday in Campinas and accepts Pix, with arrangements up to 150 reais?” will grow.
2. More visual answers, with easy-to-read tables and charts.
3. More links, from more sites. Google says it improved “query fan-out” and that search now connects users to a wider variety of sites, “including sites they might not have discovered before.”

The third point is the one that matters most to site owners. “Query fan-out” is the technical name for something simple: instead of running a single search, the AI breaks the question into several smaller searches, one for each part of the topic, and combines what it finds. Google's documentation for site owners describes it this way: the AI runs multiple related searches across different subtopics and data sources to build the answer, and so it can show a larger and more varied set of links than classic search.

Translated for your business: if the customer's question has four parts (Sunday delivery, neighborhood, payment method, price range), the AI looks up each part separately. A page that answers one of those parts well can become a link in the answer, even if the site isn't among the top results for the whole question. It's an open door for small sites, as long as the answer is written on the page.

How to get your site ready for these answers

Google repeats in its documentation that there is no special optimization for appearing in AI answers: the same SEO best practices apply, and there is no need to create a new file, special markup or an “AI file.” To be cited, the page needs to be indexed and eligible to appear in search with a text snippet. That said, what changes is where to put the effort:

• One page for each question your customer asks. Price, lead time, service area, payment method, comparison between options. If the question will be broken into parts, each part needs a findable answer. A generic “Services” page with five lines answers none of them.
• Answer in the first paragraph. “We deliver on Sundays from 8 a.m. to 1 p.m. in Campinas, Valinhos and Vinhedo” is a sentence the AI can use. “Ask about our terms” is not.
• Text, not just images. A price table in a photo and a menu in a scanned PDF are not read as text. Google asks that important content be in text form.
• Up-to-date listings. The same documentation says to check your Business Profile and Merchant Center. For local businesses, the step-by-step is in how to appear at the top of Google Maps.
• Don't block those who read your site. Google mentions robots.txt and also the services that sit in front of the site. That was exactly the problem we found on our own, which we described in your site may be blocking ChatGPT without you knowing.

The side effect: more impressions, fewer clicks

A better answer at the top of search has a cost for the site: some people read the answer and don't click. Google states that clicks coming from pages with AI answers are “higher quality,” with people spending more time on the site, but it doesn't publish how much click volume changes.

In our own Search Console, comparing the 28 days up to November 11, 2026 with the previous 28 days, impressions rose 78% and clicks rose 20%. The click-through rate fell from 1.4% to 1.0%, with the average position flat at 7.4. This can't be attributed to AI alone: part of the new impressions came from news topics, which tend to get few clicks. But it is the pattern the whole market is seeing, and we have already explained why it happens in Google pays sites used in AI answers.

The practical consequence is to measure differently. Impressions became a presence number, not a sales number. What matters is whether the people who click become contacts, and that is measured with forms, WhatsApp and phone calls, not just the click chart.

For those who use AI at work: what changes in free Gemini

For the business owner who uses Gemini as an assistant, the change is direct: the free-plan model got better. Google says Hydrogen follows the tone, format and rules you ask for more closely, so emails, texts and documents come back ready on the first try, and that it builds comparison tables without losing detail.

Three uses worth testing this week:

• Answer emails in your own writing style. Paste three of your old replies and ask it to write the next ones in the same tone.
• Compare quotes. Send the proposals from three suppliers and ask for a table with price, lead time, warranty and what each one left out.
• Summarize a Drive folder. Gemini pulls context from Gmail and Drive when you authorize it. To work on a fixed set of documents, the route we have already described is Gemini Notebook.

One caveat: Gemini inside Google Sheets with Hydrogen starts with subscribers to the paid Google AI Pro and Ultra plans and eligible business subscriptions. On the free plan, access is through the app and automatic model selection.

For those who pay for AI inside a system: the pricing math

This part matters to those who have (or want to have) automated customer service, order classification, contract summarization or any system that calls AI through the API, paying per use. Prices are in dollars, per million “tokens” (word fragments; a million is equivalent to a few hundred pages of text):

• Gemini 4 Hydrogen: US$ 1.50 for input and US$ 7.50 for output, a single price across the entire 1-million-token context.
• Gemini 4 Carbon and Argon: US$ 2 and US$ 10 until November 30, 2026; from December 1, US$ 4 and US$ 20.

Developer Simon Willison, who closely tracks model pricing, called the table “confusing”: today Hydrogen is only 25% cheaper than its bigger siblings, and the gap becomes 75% in December. He also points out that Claude Haiku 5.5 and GPT-6 Luna cost US$ 0.10 and US$ 0.50, meaning Hydrogen is 15 times more expensive than those two.

The rule of thumb that comes out of this is simple: the right model depends on the task, not the brand.

• Short, repetitive task (classifying a message, extracting data from an invoice, answering a frequent question): use the cheapest model that gets it right. At high volume, the 15-fold price difference weighs more than a few intelligence points.
• Multi-step task (reading documents, deciding, calling other tools): this is the tier of Hydrogen and its direct competitors. Test with your own cases before choosing.
• Hard, rare problem (reviewing a complex contract, designing a system): it's worth paying for the strongest model, because the volume is low.
• Before December 1: if your system uses Carbon or Argon, the bill doubles. It's time to measure how much each task consumes and see whether part of it can move to a smaller model.

Google itself points in this direction by launching a feature that switches between models within the same task: the strongest decides, the cheapest executes. In a test by the company, this came close to the strongest model's result while spending 24% less. Setting up this split in a small business system is the kind of work we do in AI implementation for companies.

What doesn't help

• Switching AI with every launch. Every month a “better” model comes out. Those who already have a working flow gain more by testing the new one on ten real cases than by migrating on announcement day.
• Writing “for the AI.” Google says in no uncertain terms that there is no special markup or file. Confusing, repetitive text, or text written to trick robots, is still bad text.
• Believing lab numbers as if they were your case. “2.5 times faster” and “five times less invention” are averages from the manufacturer's tests. Your customer service has its own questions.
• Measuring success by clicks alone. With more ready-made answers at the top, clicks fall and the contacts that arrive tend to be more qualified. Without measuring contacts, the conclusion comes out wrong.

What this means for your business

The launch looks technical, but the part that matters is distribution: Google put a new model in front of everyone who searches, for free. Search answers get longer, understand more complicated questions and cite more sites. For small businesses, this opens space, as long as the site answers the customer's questions one by one, in text, and nothing is blocking those who read it.

If you want to know how your site appears in these answers and which pages are missing to cover your customer's questions, our SEO and AI-answer optimization team does this assessment and delivers the list of what to write first.

Sources

• Google, official blog, “Introducing Gemini 4 Hydrogen,” by Koray Kavukcuoglu, November 12, 2026.
• Google Search Central, “AI features and your website” documentation.
• Simon Willison, “Release: llm-gemini 0.37,” November 12, 2026, with the Gemini 4 family pricing table.
• Artificial Analysis, X post of November 12, 2026 on the Gemini 4 Hydrogen intelligence index.
• Google Search Console for the Agência Primeira Página site, periods from September 17 to October 14 and from October 15 to November 11, 2026.