Agentic Search vs. classical search
Agentic Search vs. classic search
What Google's I/O announcement really changes
For 25 years, every Google search began the same way: a white input field, a blinking cursor, and the silent prompt to condense your question into three keywords. At the Google I/O 2026, the annual developer conference held in May, the company declared this era itself over. Replacing the search box is an expandable input area based on Gemini 3.5 Flash, which accepts text, images, files, videos, and even open Chrome tabs. Even more importantly, users can create so-called Information Agents in the future—AI agents that continuously track a topic in the background and notify themselves when there's something new.
So far, only users with a Google AI Ultra plan can enjoy these features. Starting in the summer of 2026, the broader rollout is expected to begin for more users initially in the USA.
This makes Agentic Search a product reality from a futuristic scenario at the world's largest search provider. It's time to properly categorize the concept.
What is Agentic Search?
Agentic Search refers to a search where the work is not done by a human, but by an AI agent. The agent understands a goal ("Find me a 3-bedroom apartment in Leipzig under 1,200 euros with utilities included, with a balcony, available from September"), independently breaks it into sub-steps, searches multiple sources, compares the results, checks them against each other, and delivers a refined answer in the end. If necessary, it even acts: It schedules a viewing appointment, books a table in other scenarios, or adds products to the shopping cart. Three characteristics distinguish a search agent from a regular chatbot. It operates in multiple steps instead of a single question-answer round. It makes its own decisions along the way, such as deciding which source is trustworthy and which to discard. And it can remain active for longer periods, instead of forgetting what it was about after each answer. This last point is at the heart of Google's announcement: The Information Agents continue to run around the clock, according to Google, until the task is completed. For those familiar with the term from the tech corner: Agentic Search goes beyond classic RAG (Retrieval-Augmented Generation). RAG retrieves relevant documents once and formulates an answer from them. An agent, on the other hand, decides for itself whether it needs to search again, where to search, and when the answer is good enough.
How the classic search works
To compare, it's worth looking at what Google has been doing since 1998. Crawlers read websites, an index stores their key information, and a ranking algorithm sorts them by relevance in the context of specific keyword queries. The result is a list of links. The actual intellectual work - opening, reading, comparing, evaluating sources - remains with the user. This model is remarkably robust and remains the fastest solution in many cases. Someone looking for the opening hours of the city library doesn't need an agent that researches for five minutes. The limitations become apparent with complex, multi-part queries: Planning a trip with a budget, time frame, and dietary restrictions requires dozens of individual queries and many open tabs in classic search.
The direct comparison
| Criterion | Classic Search | Agentic Search |
|---|---|---|
| Input | Keywords | Goal in natural language, also with images, files, tabs |
| Result | List of links | Prepared answer, partly with completed task |
| Who works? | The user filters and evaluates | The agent researches, compares, acts |
| Time horizon | One request, one answer | Continues for days or weeks if needed |
| Transparency | High - every source is visible | Limited - the agent selects sources itself |
| Typical case | Fact query, navigation, local search | Research, comparison, monitoring, booking |
The housing example makes the difference tangible. Traditionally, it's about scouring real estate portals, setting filters, being notified daily, and checking. With an Information Agent, you set the criteria once, and the agent notifies you as soon as a suitable listing is online - exactly this scenario is cited by Google as an application case, along with notifications of new products or developments around a topic.
What Google has specifically announced
Since Google I/O 2026, it's clear: Agentic Search is already a product and not just a vision. The key components:
Information Agents
Users create, configure, and manage multiple agents in parallel, which permanently monitor topics - such as a planned trip, a legislative process, or a product category. The agents summarize multiple sources, reconcile contradictions, and provide action recommendations. The launch is scheduled for summer 2026 initially for subscribers with the Google AI Pro and Ultra plans.
The new search interface
Instead of ten blue links, the search increasingly generates interactive answer surfaces: generative user interfaces, individual visualizations, even small mini-apps that can be created using a natural language description.
Agentic booking and Universal Cart
The search can complete tasks, not just provide information. Google is expanding this to events and local services. Additionally, the Universal Cart is introduced: a smart shopping cart that collects products from Search, Gemini, YouTube, and Gmail in one place, monitors price trends and availability, and completes the purchase optionally via Google Pay or directly on the retailer's site. The rollout begins this summer in the USA and includes Nike, Walmart, Target, Sephora, as well as Shopify merchants.
The scale behind it is impressive: AI Overviews now reach 2.5 billion users per month according to Google, the dialogical AI Mode has crossed the billion mark , and the volume of requests there has more than doubled quarter by quarter. The pressure from ChatGPT, with around 900 million weekly users (as of February 2026), likely explains the pace further.
Where Agentic Search (still) falters
As impressive as the demos at the conference were - a neutral assessment also requires looking at the other side. First, the transparency: When the agent selects sources, the user only sees the result of that selection. Errors and hallucinations are harder to spot than in a list of links where you can check each source yourself. When using answers for important decisions, you should continue to open and review the agent's source citations. Secondly, the business model of the web: If agents visit websites instead of humans, the ad-supported model of many publishers collapses - a point raised early by analyst Ben Thompson, among others. Retailers and service providers can view agents as a new sales channel; for media and content providers, the equation is significantly more uncomfortable. Thirdly, the availability: As with almost all Google AI features, the innovations start first in the USA; the Information Agents are also behind a paywall. Those planning in Germany should expect delays and not confuse announcements with the available product. And fourthly, classic search remains simply superior in many situations: for simple facts, navigating to a known site, or for quick local searches. An agent that first thinks is slower here than an index that responds immediately. Thus, no replacement is realistic, but an division of labor - and Google itself integrates both modes into one and the same product.
What this means for websites
For site operators, the task shifts. It's no longer enough to be found and clicked; a website must be selected by an agent as a reliable, clearly evaluable source. On May 15, 2026 - shortly before I/O - Google published a Search Central specific guide to optimization for generative AI features . The core message: Solid technical SEO, clean structures, and robust content remain the foundation on which agents also build. From Google's perspective, optimization for generative search is simply still SEO - AEO and GEO are expressly not treated as separate disciplines.
Notably, the myth-busting section of the guide is significant. Neither breaking down content into AI-friendly chunks ("chunking") nor rewriting texts for language models or special AI markup is considered necessary by Google; their own systems can extract the relevant passage from multi-topic pages. What counts instead are contents with their own informational value: testimonials, proprietary data, a recognizable position. Interchangeable summaries of what is already available online are sorted out by AI systems as mass-produced material — the models already know the general knowledge.
A detail is currently causing confusion: the llms.txt file. According to Google, it is not required for visibility in Google Search. However, almost simultaneously, Google's Lighthouse tool with version 13.3 (released on May 7, 2026) incorporated the new "Agentic Browsing" category into its standard audit — and checks, among other things, whether a page provides an llms.txt file. The contradiction is smaller than it seems: Search doesn't need the file; for browser agents, Chrome considers it an optional auxiliary signal. The category also checks accessibility structure, layout stability, and the new WebMCP interface, is explicitly marked as "in development," and deliberately does not assign a score from 0 to 100. So those currently being offered llms.txt packages: the file doesn't improve rankings, for agentic usage it is at most a small addition.
Even fresher is a format that has been causing discussion since June 12, 2026: the Open Knowledge Format (OKF). It was introduced by Google Cloud, meaning the data division around BigQuery, not by the search team. OKF stores knowledge as a directory of simple Markdown files, each file representing a single concept, so AI agents can target access without reading entire documents. Crucially for classification: Google describes OKF as an internal data format for corporate knowledge, explicitly not as a ranking signal and not as something to publish like a web sitemap. Parts of the SEO scene, notably Marie Haynes, are already applying the idea to websites — knowledge as a cleanly linked concept graph instead of a website clutter. This is an understandable bet on the direction in which the machine-readable web is moving. However, as of today, no agent reads an OKF file hosted on a website. Experimenting now builds experience; however, it doesn't have a visibility effect yet.
The guide becomes more specific with structured data (not very important), clear product and company information, and site architecture that machines can evaluate directly. Vendors should also keep agentic booking and the Universal Cart in mind — there it is decided whether one's shop is "operable" for agents at all. Google provides an open interface with the Universal Commerce Protocol, through which retailers can connect product data, shopping cart, and checkout to agents; those who test early gain experience before the channel gains sales relevance.
Use Agentic Search sensibly in everyday life
Whoever wants to try search agents needs to ask differently than before. Keywords are of little help to an agent; what it needs is a complete goal with criteria and context. Instead of "running shoe test," something more comprehensive — perhaps even spoken: "I run ten kilometers on asphalt three times a week, tend to overpronate, and am looking for a shoe under 150 euros — compare suitable models and name the sources." The more precise the constraints are, the less the agent has to guess. And it's in guessing that most bad answers arise. The second habit worth adopting: expand sources. Every reputable search agent links its references; ChatGPT, Perplexity, and Google's AI Mode do this by default. For answers involving money, health, or important decisions, clicking on at least two of these sources is essential. If the original source deviates from the agent's response, the source wins — not the agent. Currently, Google's AI Mode is recommended as the lowest-threshold entry, Perplexity for source-heavy research, and ChatGPT for tasks that combine research with further processing. The new Information Agents remain initially reserved for Google AI Pro and Ultra subscribers and do not start everywhere. The question remains when the search box is the better choice. Always, when the answer is a single, easily verifiable fact or the goal is a certain, already known website. A useful rule of thumb: Would the research by hand take more than ten minutes and require half a dozen open tabs, the agent is worth it. Otherwise, the traditional search is faster — and also more transparent.
Conclusion
The I/O 2026 marks the point where Agentic Search transforms from a technical term into a standard product. The classic search does not disappear because of this; it remains the faster tool for simple questions. But in all areas where research, comparison, and monitoring are required, an agent will take over in the future — and websites will no longer compete just for clicks, but for the "trust" of a machine. Those who now align their content and data structures accordingly gain an edge while most competitors still watch.
FAQ
Agentic Search is not a competitor to Google; it's a new search paradigm introduced by Google itself: An AI agent researches, compares, and acts independently, instead of providing a list of links.
RAG retrieves documents once and formulates an answer from them. A search agent plans multiple steps, decides independently on further searches, and can pursue tasks over days.
Not foreseeable. For simple fact and navigation queries, classic search is faster. Google runs both modes in parallel within a single product.
Partially. When the web search is activated, ChatGPT researches in multiple stages, combines sources, and cites them — that's agentic searching. ChatGPT search does not yet offer continuously running monitoring agents in the pattern of Google Information Agents in a comparable form.
Not for visibility in Google Search — Google's own guide from May 2026 makes that clear. Lighthouse treats the file as an optional signal for browser agents. Those who can maintain it without much effort do no harm; it's not mandatory.
No. OKF is an open Markdown format that Google Cloud uses to make internal corporate knowledge accessible to AI agents. It is not a ranking signal and not a web standard like llms.txt or schema.org — confusing these three is the most common mistake. The basic idea can be applied to websites, but as of summer 2026, no agent evaluates a publicly hosted OKF file.
According to Google, starting in summer 2026, initially for users with an AI-Pro or AI-Ultra subscription. A date for wider availability is still pending.
Share this post