Semantic Optimization for AI: Understanding the New Generation of Search Engines

You don’t need to pay for a course, send your contact details to experts who only want your data, or keep jumping from video to post trying desperately to understand how to optimize your content for AIs (LLMs).

Gemini itself, in the 2.5 Pro model, shows the step-by-step process of generating information and there you have the tips you need to look at the project you are optimizing and make adjustments, which (SURPRISE!) is what we have been saying for a long time that we need to pay attention to:

• Understanding the Domain of Knowledge or subject matter
• Make your project relevant and generate authority and reputation.

In this article I will show you the reasoning that this Google model generated and that inspired me to write this article and I will base myself on Google’s own documentation to understand the concepts of relevance, authority and reputation for the search engine.

Distilling Gemini’s thinking to understand how to “Optimize for AI”

The process below was extracted from a voice search, in the Gemini app for Android, using the experimental 2.5 pro model. In the process that I describe below, we see the steps that the model generated and what are the key points to find sources to generate the answer I need:

Now let’s see how the steps used by Gemini 2.5 Pro can help us like it (and perhaps all existing LLM models) do to organize information and answer our questions.

Identify the main need

The system seeks to understand the needs of the user, in this case, the person searching, as we do when creating content and information for our projects. It already understands, in the formatting of the question, that the user needs reliable and potentially diverse information, citing by name the need for reliable sources of information.

It is necessary to remember the information search process that is triggered by the Information Gap, described below.

We are in a situation of lack, where we lack something to move from the state we are into a new one, where the information acquired allows us to meet the initial need. There are many examples, but imagine that you have been promoted at work and need to perform follow-ups on employees who are now subordinate to you.

This new development generates the need for information on how to do this monitoring, and it triggers the search process.

So it was to be expected that Gemini would seek to identify the user’s main need.

Recognize the subject

The system, after identifying the user’s main need, seeks to recognize the subject that the user is looking for and here we have a valuable tip on a subject that I always talk about: define the domain of knowledge to which your project belongs, identify the related subjects and remain faithful to this choice.

This is important because we are dealing with algorithms, LLMs, and automated systems, all of which have difficulty understanding human language and its nuances. Here on the blog, you will find several pieces of content where I talk about ambiguity and semantics, which will help you understand the need to define the topics covered and not overly broaden the scope of your content.

Formulate search queries

This is an internal thought process that is translated into the actual use of the search tool, in this case, Google Search. In the screenshot, you can see that Gemini searches for relevant sites and searches for related terms, generating a list of these terms.

As he previously acknowledges the topic, it leads me to understand that Google’s knowledge graph may be being used here to build the search strategy for terms related to the main topic and topics related to it.

This is a strategy very similar to that used by a librarian when helping someone seeking help in the library to find a book that can answer a question they have.

Run the search (using the tool)

Then it generates the command:

Gerar o bloco tool_code com as consultas selecionadas

And you do a Google search, just like the rest of us. So when we talk about the new versions of Gemini and all the other LLMs that access the Web to improve information generation, we can say with a high degree of certainty that the strategies that contribute to a better understanding of content with the aim of indexing and retrieving it continue to be valid from this stage of an LLM’s reasoning.

Analyze the search results

Here the system simulates, based on the expected results, the search results that will probably point to various types of resources that can be organized to generate the answer to my question.

We know that the system simulates the search result, suggesting that the Google search is a simulation (based on the embeds from the LLM database?) that generates a synthetic search result, with a strong probabilistic basis (research that will probably point) to resources, which can be websites, books, videos, academic articles and other types.

The next item closes the question about the assumption of the search simulation and the synthetic result assembled through this simulation.

Synthesize the findings and structure the response

Here we have the largest volume of tips on the strategies that Gemini uses.

• Prioritize Authority: Start with the most authoritative, scientifically-oriented sources.
  • Depending on the type of search, the system is guided to select some very specific categories of sources.
• Present Information Portals: Mention reliable portals.
  • Due to the nature of my question (involving health-related topics) the system takes precautions regarding the reliability of the portals accessed, indicating that some portals are selected and marked as reliable.
• Consider Associations of {assunto}
  • Again a semantic clustering strategy, creating associations of the main subject, perhaps to expand the search or even test new approaches.
• Personalize (subtly): Where Gemini uses my profile to personalize responses for me.
  • Here’s where the functionality in paid versions of LLMs differs from the free versions. I can create a profile of myself and save it in the tool, which will always personalize the search according to what I’ve told it about myself.
• Provide additional assistance: Ask if the user has specific aspects of {assunto}.
  • Another extreme difference between information retrieval and information generation: the possibility of the system having memory and inciting a conversation that expands the subject or addresses new aspects.
• Refine the language: Ensure that the language is clear, direct (according to the user’s preference) and uses appropriate terminology in Portuguese. Use Markdown for structure (markers).
  • Another aspect of personalization is the adaptation of the language to my profile described on the platform by me and the output format organized for human reading, using a structure that makes it easier to scan the response.

In Google’s Search Quality Evaluator Guidelines document you can find out what relevance, authority and reputation mean to Google.

Based on the Search Quality Raters Guidelines, relevance, authority, and reputation play crucial roles in how Google assesses the quality of search results. These concepts are intrinsically linked to the goal of providing users with useful and trustworthy information.

Prints of the official Google documentation that guides search quality analyzers.

What does Relevance mean to Google?

Relevance is key to determining whether a search result meets a user’s needs. The guidelines emphasize that evaluators should focus on user needs and assess how useful and satisfying a result is for the specific query. A query can have multiple interpretations and user intentions, and results should be relevant to the dominant, common interpretation, or even a reasonable minority interpretation.

The “Needs Met” rating scale ranges from “Fully Meets” to “Fails to Meet”. A result that is completely off-topic for the query or addresses an unlikely interpretation of the query is rated as “Fails to Meet”. Conversely, a result that is highly useful for any dominant, common, or minor reasonable interpretation of the query/user intent is rated as “Highly Meets”.

Relevance isn’t just about keyword matching. Evaluators are instructed to use common sense and web research to understand the user’s query and intent. They should consider whether people in different locations or with different contexts might be searching for something different with the same query.4 Additionally, relevance can change over time, as the meaning of queries can evolve.

In short, relevance ensures that the results returned by Google are useful and match what the user is looking for, whether it is specific information, a website, a product or a physical location.

Prints of the official Google documentation that guides search quality analyzers.

What does Authority mean to Google?

Authority is an essential component of the E-E-A-T (Experience, Expertise, Authority and Trustworthiness) concept, which is one of the pillars of page quality assessment. Authority refers to the degree to which the content creator or website is recognized as a leading source on a given topic.

The guidelines explain that while most topics do not have a single, authoritative “official” website or content creator, when there is one, that source is often one of the most trusted. Examples include the government’s official passport renewal page or a local business’s social media profile as an authoritative source for information about current promotions.

For YMYL (“Your Money or Your Life”) topics, the authority of a site should be judged by what experts in the field have to say. Recommendations from expert sources, such as professional societies, are strong evidence of a positive reputation and, by extension, authority.

When assessing authority, raters should consider whether the content creator has the necessary knowledge or skill for the topic. Different topics require different levels and types of expertise to be considered authoritative. Authority contributes significantly to the overall trustworthiness of a page. An authoritative website or content creator is seen as a “go-to” source for information on that topic, which increases the likelihood that the content will be accurate and useful.

Prints of the official Google documentation that guides search quality analyzers.

What does Authority mean to Google?

The reputation of the site and its content creators is another crucial factor in assessing page quality. The guidelines instruct raters to research reputation using external, independent sources. This includes looking for news articles, references, expert recommendations, and other credible information written by people about the site or content creator.

Reputation research should be performed according to the topic of the page. A website may be a reliable source for one type of content (e.g. funny videos) but unreliable for another (e.g. financial information).

Customer reviews are also important for assessing the reputation of stores, companies, or any website that offers products or services. A large number of positive and detailed user reviews can be considered evidence of a positive reputation. On the other hand, credible reports of fraud or financial misconduct indicate an extremely negative reputation.

Reputation is especially important for detecting untrustworthy websites and content creators. Even if the content looks good on the surface, reputation research can expose scams, fraud, or other signs of harm. An extremely negative reputation for a website or content creator can lead to the page being ranked as “Lowest” because many users would consider the website or page to be untrustworthy.

What does this mean for the projects I optimize?

In short, reputation helps Google understand how others perceive a specific website or content, contributing to the overall assessment of the reliability and quality of the information.

Taken together, relevance, authority and reputation are key elements for search quality raters, who use these criteria to help Google refine its algorithms. By identifying relevant content from reputable, authoritative sources, Google aims to provide users with the best search experience, they say.

But one thing is certain, offering answers with accurate, reliable information that meets your needs is the way to have users satisfied with the tool.

But what I want to tell you is that it is not necessary to change the way you call SEO, nor to fall head over heels for any new thing that some guru or company tries to sell us. You need to understand how the tools work, change your thinking about outdated practices, and do a lot of testing. Maybe when the search is 100% generated by LLM some things will change completely, maybe when all that’s left are agents that talk to other agents and we just send a problem and receive the solution, we will need to completely reinvent ourselves, but for now I will keep calm, study and test a lot.