Anyone who wants to optimize for Google AI Mode needs to understand how the new Google search works. A patent provides valuable insights.
Google AI Mode represents a watershed moment for search—and for SEO. Unlike traditional search, which is primarily based on document ranking, AI Mode is a multi-stage, reasoning-based system that generates natural language responses. This process has significant implications for how online content is discovered in the future, rendering traditional SEO approaches insufficient.
A Google patent describes in detail the process from the search query to the final results in AI mode. Understanding the individual steps is important because they have implications for website optimization.
Important: The results generated are not only determined by the wording of the search query, but also by many contextual factors such as:
-
- Past interactions
- Device signals
- Personal preferences
- Location
- Information from other apps
The process is described in the Google patent as follows:
The individual steps and the resulting SEO implications are described below.
Thanks to Hanns Kronenberg and Michael King for their contributions and explanations.
Step 1: The request is received
To start, a user enters their search query. This can be done via text, voice, or even image input, making it multimodal. The query triggers a more comprehensive information gathering process, which is outlined in the following steps.
Content is evaluated not only based on the exact words in the search query. It also examines, for example, how the original query relates to other possible search query-document pairs that the system later generates during query fan-out.
Step 2: Contextual information is retrieved
The system draws on information about the user and device, including previous requests in the same session, location, behaviors associated with the account, such as those from Gmail or Maps, device signals, and persistent storage representing user state. This information helps the system understand the current request in the correct temporal and behavioral context.
What does this mean for SEO? The same query can lead to completely different search result paths for two different users. This complicates ranking tracking and emphasizes the importance of a consistent presence across different online spaces.
Step 3: The first LLM output is generated
An LLM (in the case of AI Mode, it’s Gemini 2.5) processes the search query along with the collected context. The model generates initial conclusions, including inferred user intent, ambiguity resolution, and hints for classifying the query. This step is crucial because it establishes the system’s internal understanding of what the user is trying to achieve.
What does this mean for SEO? The visibility potential of content is strongly influenced by how well it matches the user intent determined by the system.
Step 4: Automatic search queries are generated
The output of the LLM forms the basis for generating multiple search queries, which the system then executes. This process of executing multiple search queries in parallel is also known as query fan-out. These queries can include related, implicit, comparative terms, or words from the user’s previous searches. The system can also add terms from the user’s context.
What does this mean for SEO? Visibility depends on various layers of the process. If content is optimized for the original query but is irrelevant to the queries generated by the system, it may not be retrieved at all. Optimization means anticipating and covering potential queries.
Step 5: Documents are retrieved
Documents matching the search queries are retrieved from the index. This happens not only in response to the original query but for the entire range of automatically generated search queries. The system collects documents relevant to the various sub-questions. The selection of documents can be influenced by various criteria, including those that depend on the user.
Content now competes in a dense retrieval environment based on semantic similarity, not just ranking.
Step 6: The request is classified based on state data
Using the original search query, context information, generated queries, and retrieved documents, the system assigns a classification to the query. This classification determines what type of response is needed, for example, an informational, comparative, or transactional response. It also checks whether further clarification is necessary.
What does this mean for SEO? The type of answer required determines what type of content is selected and how it is summarized. Content should be structured to match the respective user intent. Content that doesn’t match the user intent can be excluded, despite its relevance.
Step 7: Specialized downstream LLMs are selected
Based on the classification determined in the previous step, the system selects from a set of specialized LLMs. These can be models specifically trained for tasks such as summarization, structured data extraction, translation, or decision support. Each of these models plays a role in transforming the retrieved documents into a useful synthesized answer. Multiple models can also be selected.
What does this mean for SEO? The LLM, which ultimately interacts with the web page content, may never see the entire document and may only process a specific paragraph or element, such as a list or table. Therefore, the format and ability of content to be broken down into separate units are becoming increasingly important. This is also referred to as “chunkability.”
Step 8: The final output is generated
The selected downstream LLMs produce the final response. It is formulated in natural language and can contain information from multiple sources and different modalities (text, video, audio).
Step 9: The response is rendered on the user’s device
The synthesized natural language response is sent to the user. Often, references or interactive elements from the retrieved corpus are added. Citing a website can drive traffic to that page.
What does this mean for SEO? Visibility in the AI ​​response sources doesn’t guarantee traffic. SEOs rely on measuring the so-called “Share of Attributed Influence Value (AIV).” A citation becomes both a tool for awareness and trust-building.
Evaluation
Websites must fulfill additional requirements if they want to be successful in Google AI Mode. This involves answering various questions that may arise from users’ search queries. In this respect, this isn’t a completely new approach, as the recommendation has always been to focus on user needs and questions when creating content.
Unfortunately, it will remain unclear which questions Google will create during query fan-out –Â at least, there are currently no plans from Google to make these search queries available.
