Google's search box is now a task dispatcher

Yesterday Google said it was the biggest change to the search box in over 25 years. That part is true, and not for the reason most of the coverage is fixating on. The interesting thing is not the new model or the conversational interface or the agentic booking. The interesting thing is what the search box has become underneath all of that.

It is no longer a query receiver. It is a task dispatcher.

I have been reading the Google blog post and the AEO guide back to back for two days. The conclusion I keep arriving at is the opposite of the headlines.

From query to task

Last month I started looking for an apartment in Berlin. Two-bedroom, under €1,500, close to a Ringbahn station so my partner can get to her studio without changing trains. Under the old model the workflow was familiar. I type "apartment Berlin Ringbahn 1500 zwei zimmer" into Google, get ten blue links, click four listing aggregators, mentally merge the inventory, repeat the search with different filters, and by Friday I have thirty tabs and a spreadsheet.

I tried the same task in AI Mode this week. I typed the same thing in plain English. The model surfaced seven candidates with commute time to the office I work from, current rent comparables for each Bezirk, and offered to keep watching the saved task. The Information Agents Google announced for AI Pro and Ultra this summer will do exactly that, sit on a saved task 24/7 and notify me when something new matches. Google's own words: "in the background, 24/7, these agents intelligently reason across information to find exactly what you need."

Four announcements ladder up to the same shift. Gemini 3.5 Flash is now the default model in AI Mode for everyone globally. Information Agents are coming to AI Pro and Ultra this summer. Agentic booking, where Google literally completes a reservation on your behalf, rolls out to everyone in the U.S. this summer for local services and experiences. And generative UI, the system that composes a custom layout for your query in real time, ships free for all users this summer too.

Queries that used to be one event are becoming sessions that span days. The session is the new query, and the search box is the dispatcher that opens it.

From pull to push

The user used to fetch. They typed, they clicked, they read. The agent now fetches on their behalf, sometimes while the user is asleep. This sounds like a small UX change. It is a foundational change to what "search volume" means as a metric.

Consider what one Information Agent watching one saved task for thirty days looks like in GSC. Zero impressions for the user. Zero clicks. The agent did the work, the user got one notification at the end. Multiply that across a billion monthly AI Mode users, Google's number, and the shape of the demand curve changes.

Google said AI Mode passed one billion monthly users one year after launch, with queries more than doubling every quarter since launch. Pause on that figure. That growth is in AI Mode usage, not in "searches the way GSC counts them". Which is why the GSC impression curves operators are looking at right now feel unfamiliar; the denominator changed underneath them.

When I look at our own dashboard across the last 90 days, AI Mode citation count grew faster than impressions did. The gap between those two curves is the story. If you are still measuring SEO performance by impression count alone, you are watching the smaller of the two numbers. We wrote about the click collapse last year; the underlying shift hasn't reversed, it has compounded.

From static page to generative UI

One sentence in the Google I/O post deserves more attention than it got. Search now builds custom layouts in real time, assembling them from components on the fly. For free users. This summer.

That sentence reframes the page as a deliverable. The user is not seeing your page. The user is seeing a dynamically composed component that pulled three sentences and one statistic from your page, plus a price from a competitor, plus a review snippet from Reddit, plus a map block from someone's Google Business Profile. The page contributed. The page did not deliver the experience.

Click-through rate as a metric becomes hard to interpret in this environment. A page can be cited five times in a generative UI surface and earn zero clicks. The same page, two months prior, would have ranked at #3 and earned the click. The page got better at its job and the metric got worse. That tension breaks dashboards before it breaks the underlying SEO logic.

The optimization target is no longer "rank for the query and earn the click". It is "be the source the layout assembler reaches for". Those are correlated, but they are not the same thing. Which inputs make you the citation source is a different question, one Google partially answered with the AEO guide it shipped this week.

From pages to entities

The AEO guide hints at something Google has been quietly working toward for years. Google has indexed URLs that contain words for 25 years. Increasingly it reasons about businesses, products, and services as entities across many sources. A page is one signal among many for the entity it represents.

Read the AEO guide's "still matters" list with this lens. Google calls out Business Profiles for local visibility and Merchant Center for ecommerce product information by name. Those are entity surfaces, not page surfaces. They are how you tell Google what your business is, not what one of your URLs says.

Now read the "you don't need" list with the same lens. LLMs.txt files. Content chunking. Special schema beyond what's already supported. Those are all page-level deliverables that fragment your entity signal across new surfaces. The guide points in the opposite direction: consolidate the entity, do not split it.

Entity clarity becomes the load-bearing SEO input. Practically:

• Same business name everywhere: homepage, GBP, Wikipedia if you're notable, citations, the schema you ship
• Same address format across all surfaces; citation consistency was always table stakes for local SEO and now it matters for AI Mode synthesis too
• Same services and categories: if your GBP says "marketing agency" but your homepage h1 says "demand generation consultancy" you are blurring your own entity in the model's eyes
• Named experts with real bios, with sameAs links to LinkedIn or X or Wikipedia, with author schema that parses cleanly
• Consistent product specs and identifiers across GBP, Merchant Center, your PDPs, and any third-party listings

None of that work is new. The weight on each item shifted. Our GBP guide and our piece on local search intent both walk this in detail; both became more useful this week, not less.

Why SEO matters more, not less

This is the part of the article I want to nail, because the doomer takes are loud and they are wrong on the evidence.

Google shipped a developer document this week titled "Apply foundational SEO best practices to generative AI search". Read that heading twice. Then read the load-bearing sentence inside the section: "The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems." Google is saying, in its own developer documentation, that the foundations did not move.

Which inputs carry more weight now:

• Structured data that parses cleanly: FAQPage, Product, LocalBusiness, Person, because the model uses it to ground entity reasoning
• Named-author expertise with verifiable bios and sameAs links; the model surfaces named sources more often than anonymous ones
• Crawlable, indexable, JavaScript-rendered content the agent can read; the AEO guide notes that browser agents may access a site by inspecting the DOM structure and the accessibility tree, which means clean semantic markup is now an agent input, not just a screen-reader input
• Entity clarity across GBP, Merchant Center, site, and citations
• Semantic HTML and good page experience, because agents parse the page the way the renderer does, not the way a user does

Which inputs carry less weight:

• Keyword density and on-page variants; the model resolves synonyms, you do not need to feed it twelve of them
• AI-spun thin content; the model has its own and does not need yours
• Link velocity tactics aimed at recency signals
• The cottage industry of AI-SEO deliverables Google just disowned: LLMs.txt files, content chunking, AI-specific writing style, inauthentic mentions across the web, special-purpose schema beyond what's already supported

That last bullet deserves a closer look. The AEO guide names five things you do not need to do, by category. If you are paying for any of them, you are paying for work Google has said adds no signal. LLMs.txt files in particular sit at the top of the list; a cottage industry sprung up over the last year selling "your site's LLMs.txt strategy" and Google's official guidance is that they add no value to generative AI search.

A dogfood note before I move on. This past week we pushed FAQ JSON-LD to three articles on this blog. The CMS escaped the <script> tag on save, so the live HTML shipped <script type="application/ld+json"> as plain text. Google read zero FAQ schema on those pages. We only caught it because we view-sourced the live HTML. Structured data that does not parse is worse than no structured data: it costs you opportunity without telling you why. The fix is one line in the publisher script. Our piece on AI Overview citations walks the diagnostic step by step.

Where the "Google Zero" headlines are right

Honest acknowledgement. Doom-mongering is off-brand for us, but pretending nothing changed is dishonest, and dishonest pieces age poorly.

The HouseFresh case is real. Gisele Navarro went on Decoder in May 2024 and walked through a roughly 65% traffic drop on a high-quality independent air-purifier review site. That happened. It is happening to other review sites. Nilay Patel's "Google Zero" framing was directionally correct for a specific cluster of content types.

Content shapes exposed to the AI Overview and AI Mode synthesis drain:

• Product comparison listicles: "best mattresses 2026", "top 10 CRM tools"
• How-to content that fits inside a 300-token AI Overview summary
• Definition-style explainers; the model has its own answer for "what is a webhook"
• Roundup posts that aggregate existing sources without adding primary research
• Thin affiliate review content that reuses manufacturer copy

Content shapes still defensible, and in some cases gaining value:

• Original reporting with named sources
• Primary research with proprietary datasets the model cannot scrape
• Named expertise on narrow topics
• Transactional and local pages tied to a real physical business
• Content with non-copyable artifacts: live tools, calculators, datasets, interactive visualizations

The "is SEO dead" question collapses if you separate by content shape. For listicle content, the easy money is gone. For original work tied to real expertise or a real business, the value went up. The portfolio reading is what matters; the binary is a category error.

What an operator can control this week

Concrete. Five items. No theory.

Audit your AI-citation eligibility. Sample thirty of your most important URLs. Run each target keyword through AI Mode. Note which pages got cited and which got skipped. That gives you a baseline you can compare against in 30 days. Our checker at /tools/ai-visibility-checker/ runs this without a login.

Verify your structured data renders as executable script. View the live source on any page that should ship FAQ or Product schema. Search for <script type="application/ld+json". If it is there, your CMS is silently stripping the schema. We had this exact bug on three of our own articles last week. Our audit tool flags it on the site you point it at.

Name your authors and ship author schema that parses. Real bios, real photos, real sameAs links to LinkedIn and X. The AEO guide is explicit that expertise signals still feed the ranking and quality systems generative AI features sit on top of. Anonymous content is a citation handicap, not a neutral choice.

Tighten entity clarity. One audit pass: business name, address, services, categories, same string across GBP, your homepage, your About page, your schema, your citations. Inconsistencies blur the entity in the model's eyes. The GBP guide walks the full checklist.

Stop spending on the cottage industry Google just disowned. LLMs.txt files, content chunking, AI-specific rewriting services, inauthentic mention packages. Google named all five categories in the AEO guide. If you have an open invoice for any of them, that is the easiest budget reallocation you will make this quarter.

Sequence matters. The audit comes before the spend. If you are also re-evaluating paid spend, the calculus there changed less than the calculus on organic, though it is moving too. Founders running the program themselves can stretch this checklist across two afternoons.

Frequently asked questions

Is SEO dead now that Google has AI Mode?

No. Google's own AEO guide, published this week alongside the AI Mode announcement, opens with a section titled "Apply foundational SEO best practices to generative AI search" and states that "the best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems." What changed is the mix. Page-level click optimization carries less weight. Entity clarity, structured data that parses, named expertise, GBP completeness, and crawlability carry more. Content shapes that depended on AI-summarizable listicle traffic are taking real hits; original work, primary research, and transactional pages tied to a real business are gaining value. The portfolio reading is what matters.

What does Google's new AEO guide say I should do?

The guide reduces to two lists. Do these: ensure indexability and crawlability, follow JavaScript SEO best practices, use semantic HTML, maintain good page experience, reduce duplicate content, complete your Google Business Profile if you serve a local market, and use Merchant Center feeds if you sell products online. Do not do these: build LLMs.txt files, chunk your content into AI-friendly fragments, rewrite content specifically for AI systems, chase inauthentic mentions across the web, or overfocus on special schema beyond what Google already supports. The whole guide is at developers.google.com/search/docs/fundamentals/ai-optimization-guide and reads in about ten minutes.

Should I create an LLMs.txt file?

No. Google's AEO guide places LLMs.txt files at the top of its "you do not need to do" list. The format was promoted by a cottage industry of AI-SEO services through 2024 and 2025 with the implied promise that Google or other AI systems would prefer sites that ship one. The official guidance is that LLMs.txt files add no signal to generative AI search on Google. If you are paying a vendor to generate and maintain one for you, that line item can come out of the budget. Other AI search providers may behave differently, but for Google's surfaces the answer is no.

How is AI Mode different from AI Overviews?

AI Overviews are the AI-generated summary block that appears at the top of the traditional ten-blue-links results page for some queries. They sit above traditional results. AI Mode is a separate surface, a full conversational interface where the user types in plain English, gets a synthesized answer, asks follow-ups, and now uses agentic features like 24/7 Information Agents and booking. Both are powered by Gemini, both can cite your pages, and both can earn zero clicks for cited sources. AI Mode is the surface Google is positioning as the future of the search box; AI Overviews are the bridge from the old experience to the new one. Operators should track citation eligibility on both.

Will Google AI Mode have ads?

Not yet, but ad-supported AI is inevitable given Google's revenue model. Search ads were 57% of Alphabet's revenue in the most recent reporting period; the company cannot leave a billion monthly AI Mode users unmonetized indefinitely. The product team has signaled testing of ad-like placements in AI Overviews, and ad units in AI Mode are widely expected to follow once the surface stabilizes. The honest answer for operators planning a 12-to-24-month roadmap: assume AI Mode will look more like a paid-and-organic hybrid surface by 2027, similar to how the SERP evolved over the last two decades.