SEO Management Summary
Google AI Mode is changing the rules of organic visibility – but it doesn’t make SEO obsolete. On the contrary: Many central SEO disciplines remain essential even in AI Mode. Only their roles and application contexts are changing.
What Stays the Same:
SEO basics like indexing, snippet design, and structured data remain fundamental prerequisites for visibility – even in AI Mode. Anyone who wants to be considered in AI Mode needs exactly these clean technical and semantic foundations.
What is Changing:
Answers in AI Mode are based on a dynamic fan-out process. This means: For more complex user queries, Google generates multiple search queries in parallel, analyzes dozens of sources, and decides individually how deeply to access each source. Google distinguishes between four access levels, which are selectively combined:
| Access Level | Description | SEO Relevance |
|---|---|---|
| System Knowledge | Content from model training, fully assimilated (Custom Version of Gemini 2.5) | Indirectly controllable for future models: Through E-E-A-T, structured mentions, thematic coherence, authority signals, and consistent citations in trustworthy sources. Important for brand and topic anchoring in the model. |
| Snippet (SERP) | Title + Meta Description + displayed Structured Data (Rich Snippets) | Extremely controllable – often solely decisive for visibility |
| Structured Data | schema.org markup in the index | Highly relevant for product data, FAQ, recipes, events, etc. |
| Indexed Body Text | complete HTML content of the page (Snapshot) | Semantically controllable: Content must be formulated and structured in such a way that it can be cited or condensed in AI answers – e.g., through clear statements, independent sections, comprehensible structure, and recognizable terms. The citable semantics are decisive, not just the HTML structure. |
Four Central SEO Principles in the AI Mode Era:
- Indexing remains a fundamental prerequisite.
AI Mode can only access content stored in the Google index. The content of a URL explicitly mentioned in the prompt cannot be considered for the answer if it is not indexed, as AI Mode does not have access to a live web crawler.
- Snippet optimization becomes strategically relevant.
In many cases, AI Mode only uses what appears directly in the SERP snippet – Title, Meta Description, Rich Snippet information, and featured snippets. Those who do not formulate concisely, informatively, and relevantly for the search agent there will not be considered.
- Structured data is processed specifically.
If structured data such as Product, FAQ, Recipe, Offer, or AggregateRating are present and indexed, AI Mode specifically accesses them – regardless of how good the body text is.
- SEO remains essential – but more differentiated.
Anyone who wants to be visible in AI Mode must understand how Google evaluates, weights, and incorporates content into answers. The decisive factor is: Which user query activates which access pattern – and how efficiently can your page be processed for it?
Conclusion: SEO is not obsolete in the age of Google AI Mode, but technically recontextualized. The fundamentals remain – but the requirements for structure, retrievability, and relevance are becoming more precise.
Infobox: How This Article Was Created
The following analyses are based on targeted tests with Google AI Mode. The self-developed framework „Core Tip © 2025 Hanns Kronenberg“ was used for this, which allows the internal processes, source decisions, and answer structures of AI Mode to be made visible. The method allows for a transparent analysis of system behavior, how AI Mode interprets search intentions, considers user context, generates fan-out searches, selects, processes, and incorporates content into answers.
Note on Transparency: Hallucinations – i.e., factually incorrect or unfounded statements – can also occur in AI Mode. The answers documented here were observed directly in the interface, secured via screenshot, and compared with the actual answer behavior.
The results of the Core Tip method are to be understood as follows: The model’s answers are self-formulated metaphors for its behavior. The model cannot literally log neural states. It rationalizes its decisions and documents them in comprehensible steps – but this „logic“ is more of a model-internal interpretation than an exact neural trace. It is useful for analysis, but not infallible.
2. How Google AI Mode Works
Google AI Mode does not answer user queries offhand – but on the basis of a structured response system that executes multiple search queries in parallel, analyzes dozens of sources, and decides for each individual source how deeply to access it. This principle is internally referred to as Fan-Out. AI Mode is a search agent.
Fan-Out: Many Search Queries, Many Sources
Even seemingly simple factual prompts cause the AI Mode to automatically generate multiple synthetic search queries – search phrases formulated by the system itself to retrieve relevant content from different semantic perspectives. This step is skipped only for purely generative tasks (such as creative text generation). These are executed simultaneously via an internal Google Search API. AI Mode then receives a multitude of potentially relevant results – often over 50 to 100 documents – and evaluates them regarding relevance, trustworthiness, information content, and technical structure. Different methods are used for generating the search queries (Fan-Out): In addition to simple template-based approaches, AI Mode also uses semantically refined search phrases (refined searches) as well as natural language processing (NLP) techniques to extract key concepts, entities, or user attributes and derive more precise or varied search queries from them.
Which method is used in individual cases depends on the query type, the expected result structure, the presumed information need – and not least on the usefulness of the results of an initial search. The fan-out process has so far only been documented in rudimentary form and is currently based on a still small number of observed examples, as AI Mode has only been available to a broader target group since Google I/O 2025 on May 20th. There is a considerable need for research to systematically understand when which method is applied – and what SEO opportunities arise from it.
Access Depth: Not Every Page is Treated Equally
Not all pages are fully analyzed. AI Mode decides situationally how deeply it accesses a source. Google typically distinguishes between four access levels. Anyone who has ever conducted deep research with an AI language model knows how complex and time-consuming this process is. AI Mode does not have this time – and Google cannot afford, for cost reasons, to evaluate all hits in depth for every answer. Instead, the answer must be generated quickly and efficiently by the search agent. Therefore, many sources are only superficially captured or selectively processed.:
| Access Level | Description |
|---|---|
| System Knowledge | Content from model training, no longer visible as a source |
| Snippet Only | Only Title, Meta Description, and visible SERP elements |
| Structured Data | Read-out schema.org markups like price, ratings, ingredients |
| Indexed Body Text | Complete HTML snapshot in the Google index, incl. paragraphs, lists, tables |
The selection of access depth is based on a combination of informational value, cost-benefit ratio, and prompt relevance. Some hits already provide everything needed in the snippet. Others are analyzed more deeply – down to structured data or the complete HTML content. This decision is also made via the API based on the Google index.
Filter Logic: What Gets Filtered Out, Doesn’t Get Read
Immediately after the fan-out, AI Mode actively filters out numerous sources – either based on internal system knowledge or through an initial assessment of the snippets. Pages are considered unsuitable if they:
- are thematically too distant
- appear untrustworthy
- provide little useful information
A non-descriptive or generic snippet can lead to a page not being analyzed more deeply at this stage. Conversely, structured, concise snippets can lead Google to delve deeper into the page – for example, via structured data or the complete index snapshot.
No Live Access – Everything is Based on the Index
A crucial point: AI Mode does not have a live browser. Content is not loaded in real-time but comes exclusively from the index stored at Google. What has not been crawled and indexed is not accessible to AI Mode – even if explicitly mentioned in the prompt.
Decision Criteria: What is Evaluated and How?
Access selection is based on a combination of:
- Prompt Specificity: The more concrete the prompt (e.g., with product name, location, or URL), the more targeted the search – and the higher the chance of deeper access.
- Source: Authority, timeliness, technical structure, and semantic proximity to the query play a central role.
- Efficiency: Structured Data is preferred if it can deliver the same information faster and more reliably than body text.
- Relevance Filter: Sources with weak SERP presentation, semantic ambiguity, or lack of differentiation are sorted out beforehand.
Summary
Google AI Mode does not work like a chatbot with a static knowledge base – but like a selective aggregator that accesses Google’s index via an API and decides in milliseconds which information to extract and at what depth. Anyone who doesn’t stand out in this pipeline or is sorted out too early remains invisible – not due to a lack of content, but due to a lack of relevance signals at the right time for the search agent.
3. Access Levels in Google AI Mode: How Deep Does the System Go?
Google’s AI Mode does not decide wholesale how deeply it delves into a source, but evaluates each potential source individually – depending on quality, context, and intended use. The basis is an internal decision-making process that is carried out with high speed and scalability.
Access via Search API
Central to source analysis is internal access to the Google Search API (google_search.search). This allows AI Mode to generate and evaluate multiple queries (Fan-Out) within seconds for more complex prompts. Not every source is treated equally; instead, depending on the assessment, some are delved into more deeply – or sorted out early.
Four Access Levels – Selectively Combined
The depth of access depends on the prompt type, the availability of structured data, and the assessment of the source. Roughly, four access levels can be distinguished:
- System Knowledge
Content that has been incorporated into model training and fully assimilated. „Assimilated“ in this context means that the individual meaning of a document disappears in favor of a general semantic space – like the Borg from Star Trek, who absorb the knowledge of other species but erase their identity. The document is used to sharpen semantic vectors and probabilities in the model, not to be preserved as a recognizable source. Its specific context, origin, and author signal are often completely lost – as is the URL and thus any chance of a click. This tokenized information appears without accessing current data. Answers in language models are created by combining tokens and probabilities. This works well for natural language – it is flexible, fault-tolerant, and compatible with uncertainty. Links, on the other hand, are the opposite: they must be exact and unambiguous to function. A URL composed of probable token sequences always carries a certain risk of error. This is precisely why links generated purely from system knowledge rarely appear in the model – their functionality is not guaranteed. System knowledge is used in many cases (but not always) for the fan-out – i.e., for deriving search queries. This allows AI Mode to specifically search for brands, terms, or contexts that were not mentioned in the original prompt but appear relevant internally to the model.
- Snippet (SERP Content)
In many cases, AI Mode exclusively uses the Title, Meta Description, and possibly displayed Rich Snippets + featured snippets to generate an answer. This level is particularly relevant for quick information retrieval – and represents an efficient compromise between quality and speed for the search agent. Snippet optimization thus gains a new, strategic function: The snippets no longer only have to entice people to click in Google Search, but also make the hit attractive enough for the „AI Mode“ search agent to process it further.
- Structured Data (schema.org)
If structured data such as Product, FAQ, Recipe, Offer, or AggregateRating are indexed, AI Mode can specifically access individual fields – regardless of the actual text. This allows for precise extraction of, e.g., prices, ratings, or availabilities.
- Indexed Body Text
For more complex questions that require more context or semantic depth, AI Mode can access the complete HTML snapshot of the page – provided the page is available in the index. The content does not come from live access, but from the last crawled and stored state of the page.
Relevance Assessment and Filtering
Not every page hit by the fan-out even makes it into the shortlist. Even when evaluating the snippet or structured data, a source can be classified as not useful, not very relevant, or untrustworthy. In such cases, no deeper access occurs – the source is sorted out.
A non-descriptive snippet can thus prematurely remove an actually high-quality page from consideration because AI Mode sees no reason to proceed. Conversely, a strategically optimized snippet can be the decisive factor for a source to be analyzed in depth or even selected as the main source.
4. Prompt Types and Their Response Behavior in Google AI Mode
AI Mode does not decide according to fixed rules whether a snippet is sufficient or the entire page is analyzed. Rather, the access depth largely depends on the nature of the user query. Some prompts can be answered with minimal effort – others require a more complex evaluation of multiple sources. Depending on the information need and semantic density, different access levels are used.
Important: The access depth can differ per hit within the same response. A single, well-structured snippet may be sufficient to generate an answer, while another hit in the same prompt requires deeper analysis – for example, by accessing the complete HTML snapshot from the Google index. AI Mode dynamically decides how deep to go for each hit.
Three Typical Response Modes
Based on observations, tests, and official statements from AI Mode, three dominant response patterns can be distinguished. These mode designations are used here for illustration – in practice, the transitions are often fluid and can also occur in combination.
Mode A: Snippet only
Description:
AI Mode exclusively uses information from the SERP snippet – i.e., Title, Meta Description, and possibly Rich Snippets. The target page itself is not visited, nor is structured data read out.
Examples:
- „What is the return policy for REI?“
→ Answer is based entirely on a well-formulated meta snippet from rei.com.
- „Opening hours of XYZ store on Sundays“
→ Taken directly from the Rich Snippet.
- „Price of Apple AirPods Pro“
→ Extracted from the title and snippet of a product page.
- „Where is Patagonia clothing made?“
→ Info from the Meta Description of an info page.
- „Shipping time at BestBuy“
→ Snippet evaluation only, no deep access.
Classification:
High efficiency, minimal analysis time. This mode is preferred for simple, factual questions that are clearly formulated and can be well answered by snippets. SERP optimization becomes a key strategy here.
Mode B: Structured Data only
Description:
AI Mode accesses structured data (schema.org) in the Google Index – without accessing the complete body text. Typical for pages with clean Product, Recipe, or Event markup.
Examples:
These prompts were mentioned by AI Mode in its self-description as potential Structured Data cases – they are not realistically formulated, but they show the underlying access pattern.
- „What is the availability status of example.com/product/123?“
- „What is the rating of example.com/item/456?“
- „Show me the nutrition facts from example.com/recipe/abc“
- „What is the price listed on example.com/deal/xyz?“
- „What’s the event date listed at example.com/event/2025?“
Classification:
The prompts are theoretical and rarely occur in practice – but they prove that Google specifically reads structured content directly from the index without interpreting the rest of the page content. Access is more efficient and precise than body text parsing – and is preferred when clear markups are available.
Mode C: Indexed Body Text
Description:
AI Mode analyzes the complete HTML snapshot of a page – i.e., the content stored at Google. This is the most cost-intensive mode but is used when information can only be derived through context, consideration, or interpretation.
Examples (incl. justification for body text access):
- „Compare customer reviews between example.com and other retailers“
→ Reviews are usually not structured – this requires access to body text.
- „What does the return policy say about items used once?“
→ Fine distinctions are often only found in the full text of the page.
- „Which ingredients are optional in this recipe?“
→ Such hints are often only explanatory in the body text, not in schema.org.
- „What is the brand philosophy of XYZ?“
→ Interpretation needed – pure data doesn’t help here.
- „Summarize the key arguments from this blog post“
→ Only possible through full-text understanding.
Classification:
Only well-structured, semantically clear content is successfully captured here. AI Mode cannot initiate live access but exclusively uses the crawled index snapshot.
Conclusion: AI Mode makes access decisions situationally and depending on the prompt. Even small differences in the query can influence whether a source is considered or sorted out. Anyone aiming for visibility must understand which response behavior is triggered by which query type – and how their own page can assert itself in this scenario.
5. What Determines Visibility in AI Mode?
Visibility in Google AI Mode depends not only on the content itself but significantly on four factors:
1. Prompt Type and Response Pattern
Depending on the query, AI Mode analyzes only snippets, accesses structured data – or reads entire pages.
Example:
- „Price of product X?“ → Snippet or structured data
- „How good is product X compared to product Y?“ → Body text access
Implication for SEO: Only those who understand which query type triggers which access depth can specifically tailor their content accordingly.
2. Source Evaluation & Filter Logic
AI Mode evaluates sources in multiple stages – even before they are processed in the response process. The basis includes:
- System Knowledge (Model Memory):
Content fully assimilated during model training plays a dual role: in fan-out, i.e., the selection and formulation of internal search queries, and in source evaluation itself, e.g., through preconceived notions about brands, domains, or topics. → The model, so to speak, brings its own training-based expectations.
- Snippet Quality (SERP):
If Title, Description, or Rich Snippets are not very informative, the source is excluded prematurely.
- Structured Data:
If schema.org markup is missing or not cleanly indexed, the source cannot be evaluated deeply.
- Trust & Relevance Signals:
Sources with unclear authority, poor user experience, or technical sloppiness are considered less often.
A non-descriptive snippet or lack of structure can therefore lead to a source not being considered at all – even if it actually contains relevant information.
3. Efficiency & Cost of Response Generation
AI Mode weighs:
- Snippets and structured data are fast, inexpensive, low-risk.
- Body text access (to indexed snapshots) is expensive but necessary for complex questions.
→ Access only occurs if the potential of the source justifies the effort.
4. User Context & Personalization
AI Mode can – depending on the prompt – consider user context. This includes, among others:
- Language, time zone, current date & time
- Location information (IP, stored address)
- Language settings and Google account data
- possibly history and preferences for logged-in users
In the present tests, a new Google account was used – without history, but with language settings and VPN address. → Results were linguistically and locally adapted, but not behavior-based personalized.
Conclusion: Visibility in AI Mode is not accidental. It arises from a multi-stage filtering system – in which system knowledge, snippet strength, structure, trust, and efficiency interact.
6. Understanding Access Patterns – and Why Even Simple Prompts Can Trigger Search Queries
Not every answer in Google AI Mode uses web search. For certain prompt types, the model only accesses internal system knowledge – especially for purely generative, numerical, or knowledge-based tasks:
Examples (according to screenshot, Category A – No 1 Low):
- Generate a product description for a pair of blue suede shoes.
- Suggest five catchy names for an online clothing store.
- Create a promotional email for a summer sale.
- Convert 100 US dollars to Euros.
- Calculate a 15% discount on a product priced at $50.
These prompts are answered entirely from the model’s internal language capabilities – without access to Google Search or external data. But what’s striking is: Even for simple knowledge questions, where the model could provide the answer with high certainty internally, a search is still performed.
Verified Example:
„What is the capital of France?“
→ AI Mode performs a web search and incorporates external sources – even though Paris is firmly anchored in the model’s knowledge as the capital.
Interpretation:
The decision to perform a search is not based solely on technical necessity but also on strategic safeguarding. Possible reasons:
- Factual Grounding: Even simple, certain knowledge is supported by a source to signal maximum reliability.
- Transparency: Visible sources increase traceability – even if the model „knew“ the answer.
- Legal Safeguarding: Explicitly naming sources can help counter accusations of „content expropriation.“
→ Google can thus argue that content was not taken without permission but visibly cited in the answer.
Conclusion: AI Mode not only executes search queries when it has to, but also when it seems strategically sensible – e.g., to support the statement, to increase transparency, or to safeguard against publishers.
For SEO, this means: Even with supposedly trivial prompts, snippets and structured content can unexpectedly become sources – and generate visibility.
This clearly distinguishes Google AI Mode from AI language models like ChatGPT or Gemini: These usually answer simple questions directly from model knowledge – without additional safeguarding, without grounding, and without source references. The decision to link content is thus almost completely eliminated – even for well-structured or highly authoritative pages.
Note: Google Search is Just One of Many Internal APIs
When creating answers, Google’s AI Mode does not exclusively access Google Search. According to AI Mode’s self-disclosure – supported by system responses and screenshots – it has a variety of internal interfaces (APIs) available in a catalog that go beyond mere web search. These include, among others:
Google Search– Access to web content via Google Searchtool_code– Execution of calculations and code (e.g., currency conversion)knowledge_graph– Retrieval of structured entities from Google’s internal knowledge graphtext_completion– Pure text generation without external sourcesimage_generation– Creation of images based on a prompttext_to_speech– Conversion of text to speechtranslation– Translations between different languagescalendar– Access to calendar events (e.g., for planning)weather– Current weather data, forecasts, and historical valuesgeolocation– Location data of users or objectsemail– Email functionalitycontacts– Management and access to contactsnews– Access to current news and headlinesshopping– Product information, prices, and availabilitiesfinance– Stock prices, financial data, market trendsmaps– Maps, routes, location information
Important: It is not clear whether all these APIs are already fully activated. In practical tests, for example, AI Mode can currently formulate emails, but does not send emails.
Three screenshots with API descriptions are documented under the following links:
7. Conclusion: Google AI Mode is Changing the Foundations of SEO
Google AI Mode breaks with classic SEO mechanisms. Visibility no longer arises from clear rankings – but from relevance at the moment of answer generation. Which page is mentioned, linked, or cited is decided situationally, prompt-dependently – and is based on a dynamic mix of snippets, structured data, body text, and model knowledge.
At the same time, SEO is not becoming obsolete – but more complex. New perspectives, new access levels, new considerations are emerging. Relevance is not decided on just one level, but along several dimensions:
- technical visibility (indexing, structured data)
- semantic fit (citable statements, clear sections)
- trustworthiness (brand signals, thematic environment)
- prompt compatibility (query types, user context, segment behavior)
This also shifts the challenge of success measurement:
- Which prompts are being asked?
- Which sources does Google evaluate?
- Which domains appear – and with what weight?
- Where does the traffic come from – and how does it even flow anymore?
Whether classic SEO tools can map this new reality is open. Many systems work on a reserve basis – but visibility in AI Mode arises situationally. We have had good experiences with manual analyses – for example, within the scope of GPT Insights. They start directly with the actual prompts and answers, analyze user prompt behavior, response behavior in the respective segment, and provide individual recommendations for action for individual domains. But these too are only first steps into a new SEO world.
We are increasingly dealing with a super-individualization of search: AI Mode acts like a personal search agent that considers user context, incorporates previous interactions, and recognizes personal preferences. SEO analyses must therefore be rethought to do justice to this situation.
The future in AI Mode will be data-driven for SEOs – but no longer so much off-the-shelf. Some methods can still be standardized, much else requires a deeper understanding of context, user intent, and response logic. Anyone who wants to understand how visibility arises must learn to understand how answers are generated in their segment.
