Google Search Is Now Written by Gemini. The Blue Links Are Not Dead, but the Business Model Is.

At Google I/O 2026, the company said it made Gemini 3.5 Flash the new default model in AI Mode, globally, and folded AI Overviews and AI Mode into what it calls “one, seamless AI Search experience.”^[1] It also announced the biggest change to the search box in over 25 years, and put AI Mode past one billion monthly users.^[1] A lot of coverage read that as: the ten blue links are gone, Gemini writes every result page now, the web is over.

That is not quite what happened, and the distinction matters, so let me be precise before I make the argument.

Google's own description of the new flow is that you go “effortlessly from your question, to a search results page with an AI Overview, to a follow-up in AI Mode.”^[1] A search results page still exists. It still has links on it. AI Overviews sit on top of it. AI Mode, the fully conversational surface where Gemini writes prose and drops citations inside the answer, is the default entry point rather than the entire product. The classic web results are still there. They are one tab away, below the fold, or one scroll down. So no, Google did not delete the ten blue links.

It did something more effective. It moved them to where nobody looks.

The click math is the whole story

Pew Research Center did the cleanest study I have seen on this. They tracked the actual browsing behavior of 900 US adults across 68,879 Google searches in March 2025, back when AI summaries appeared on only about 18% of searches.^[3] The results were not subtle.

Read the second number again. One percent. Google's answer to “you are stealing publisher traffic” has always been “we cite the sources, the links are right there.” The links are right there. Users click them one percent of the time.^[3] A citation is not a referral. It is an attribution footnote, and attribution does not pay for a newsroom.

The zero-click number is the trend line that should scare you. SparkToro, using Similarweb clickstream data, put US zero-click searches at 68.01% for January through April 2026, up from 60.45% in 2024 and roughly 45% a decade ago.^[4] That is not an AI-era anomaly. Google has been slowly eating its own referral function for ten years, with knowledge panels, featured snippets, and its own properties. Gemini just took the last third in about eighteen months.

How the answer actually gets built

Worth understanding the mechanics, because the mechanics explain why the click disappears and why some kinds of content survive.

Google calls its retrieval approach “query fan-out.”^[5] Your one question gets decomposed into many sub-queries, each of which hits the search index. The retrieved passages get fed to the model as grounding (the technical term for stuffing real source text into the prompt so the model writes from documents instead of from memory). Gemini writes prose over those passages. Then a citation layer maps sentences back to the URLs they came from.

The key structural fact: the unit of value is now the passage, not the page. Google crawls your page, extracts three sentences, and those three sentences become part of a synthesized answer that is complete on its own. The user got what they wanted. There is no unmet need left to drive a click. That is not a bug in the design. That is the design working.

And there is no opt-in and no opt-out that makes economic sense. Google's own documentation says there are “no additional requirements to appear in AI Overviews or AI Mode” beyond normal indexing.^[5] You can block your content from being used in a generated answer with nosnippet, max-snippet, or noindex, but those are the same directives that control whether you appear in Search at all.^[5]There is no lever that says “rank me but do not summarize me.” Penske Media, which owns Rolling Stone, Billboard, and Variety, made exactly this the core of its federal antitrust suit against Google in September 2025: publishers are forced to choose between letting Google synthesize their journalism or being deranked out of existence.^[9,10] Penske says its affiliate revenue fell by more than a third.^[9] Google calls the claims meritless.^[10]

Google is completely fine. That is the problem.

Here is the part people get wrong when they doom-post about this. The collapse of the referral click is not hurting Google. Alphabet did $109.9 billion in Q1 2026 revenue, up 22% year over year, with Search and other advertising up 19%.^[7,8] Pichai said queries are at an all-time high, that Search latency is down more than 35% over five years, and that the cost of core AI responses is down more than 30%.^[7]

Think about what that combination means. Query volume up. Ad revenue up. Cost per generated answer down. Clicks out to the web down by half. Google found a way to keep the demand side of a two-sided marketplace while quietly cutting the payment to the supply side.

Google's defense is that the clicks that do happen are better ones. The Search Central documentation says that when people click from a results page with an AI Overview, “these clicks are higher quality,” meaning users spend more time on the site.^[5] I believe that. It is also completely beside the point. If you cut the volume in half and raise the quality by 20%, the publisher whose business was priced on volume is still dead. “Fewer but better visitors” is what you say to a business you are shrinking.

Who actually gets hurt, and in what order

Not everyone loses equally, and the ordering tells you what to do.

• Affiliate and review sites die first.If your entire business is ranking for “best noise cancelling headphones” and collecting a commission, you are the most compressible thing on the internet. Your content is a comparison table. A model can rebuild that table from ten of your competitors in 200 milliseconds. Penske's affiliate revenue falling more than a third is the canary.^[9]
• Ad-supported informational publishers next. Recipes, health explainers, definitions, how-to. This is the category where the answer is short, the page is long, and the page is long only because ads need scroll depth. The AI answer is strictly better for the user. That is uncomfortable to say and it is true.
• SaaS content marketing loses its arbitrage.The whole B2B playbook of the last decade was: write 400 blog posts targeting “what is [thing]”, rank, capture email, nurture. That funnel was built on top-of-funnel informational search. Top-of-funnel informational search is precisely what generated answers absorb. You can still write the 400 posts. They will feed Gemini's answer and convert nobody.
• Hard news is hit differently. Google is more careful about generating on breaking news, so the summary hit is smaller. But news publishers were already losing the referral game to social and to the general decay of the open web. AI search accelerates a curve that was already bending down.

What still gets a click

This is the part worth being concrete about, because it is where the money goes next. The click survives when the answer cannot substitute for the destination.

1. Transactional intent.A model can tell you which running shoe is best. It cannot put the shoe in your cart. When someone wants to buy, book, sign up, or download, they have to arrive somewhere. This is also, not coincidentally, exactly where Google's ad inventory lives, which is why Search ad revenue is up 19% while publisher clicks are down.^[7]
2. Primary sources and proprietary data.If you have the filing, the dataset, the benchmark, the court docket, the internal numbers, the AI answer has to point at you and the serious reader has to come look. Pew's own study is a good example. You cannot summarize your way past the actual methodology.
3. Community and identity.Reddit, Discord, forums, group chats. People go there for other people, not for information. That is not summarizable, and it is why Reddit's data was worth licensing in the first place.
4. Brand and direct. The most durable channel left is someone typing your name. Every newsletter, podcast, YouTube channel, and community that owns its audience directly just got structurally more valuable, because it does not route through a ranking algorithm at all.
5. Tools and interactivity. A calculator, a live dashboard, a playground, a product. Text can be summarized. A running application cannot.

So what is SEO now?

A lot of people are selling “GEO” and “AEO” courses right now. Mostly it is repackaged SEO with a new acronym. But the job did genuinely change, and here is my honest read of what it became.

Ranking was a competition to be the destination. Being cited is a competition to be the source material. Those optimize differently. Retrieval operates on passages, so structure matters more than page-level keyword density. Be the one who states the fact cleanly, with a number, in one paragraph, in a way a model can lift and attribute. Be the origin of a claim rather than the 400th restatement of it, because grounding pipelines do dedupe and they do prefer sources they can rely on.

And accept that being cited is now a brand impression, not a traffic channel. That is the mental reframe that actually helps. If your name appears in the answer, treat it the way you treat a mention on a podcast. It builds recognition. It does not build a session. Budget accordingly, which for most companies means moving money out of volume content and into things that create direct relationships. If you are still measuring your content team on organic sessions in 2026, you are managing to a metric the platform decided to stop producing.

My actual position

I do not think this is the end of the web, and I am tired of that take. The web survived portals, it survived Facebook, it survived the app store era. It will survive this.

What I do think is that a specific, large, and mostly artificial economy is ending, and its end was overdetermined. An enormous share of the internet existed to intercept a search query and show you an ad on the way to an answer. That was never a great business for readers. It was a toll booth built on Google's decision to send you through a middleman rather than answer you directly. Google has now decided to answer you directly. The toll booth was always Google's to remove.

The genuinely uncomfortable part is not that the middlemen are dying. It is that the good stuff was cross-subsidized by the middlemen. Real reporting, real research, real technical writing were funded out of the same ad pool as the recipe blogs. Google is taking the whole pool and keeping the ad revenue, and the mechanism by which it now pays for the content it consumes is a footnote that gets clicked one percent of the time.^[3] That is not a market. That is extraction with a citation attached.

My bet, and this is where I will put a stake in the ground: the next five years are very good for anyone who owns a direct audience, anyone with proprietary data, anyone building an actual product, and very bad for anyone whose asset is a library of text that ranked. Content licensing deals will get signed and they will not come close to replacing the ad revenue, because Google has no incentive to pay market rate for something it can already take. The regulatory fight, which Penske started and others will join, is the only thing that changes the price.^[9]

And if you are a builder reading this, the practical conclusion is short. Stop optimizing to be found. Start building things people come back to on purpose. That was always the better business. Search traffic just made it possible to avoid for twenty years, and now it doesn't. Same lesson the distribution-versus-model fight teaches everywhere else in AI right now: own the relationship, or rent it from someone who does.