Based on the official Google Search Central documentation on how to optimize a website for views created with AI and AI mode.
Google says, bluntly, that optimizing for search with generative AI is still SEO. The views created with AI and the AI mode do not work as a separate channel with its own rules. They rely on the same positioning and quality systems that already decide who appears in the classic results.
Google’s guide boils down to five ideas:
- The basics of SEO still rule: clean technical structure and unique content.
- Non-generic, useful and made-for-people content is what weighs the most, more than any other recommendation.
- The “special” tactics of AEO and GEO (llms.txt, fragmenting content, rewriting for AI, fake mentions) are not needed for Google.
- Performance is measured in Search Console, in the generative AI performance report.
- Experiences with agents, which are still at an early stage, should be monitored.
If you only have time to read this section, you already have the essentials. The rest of the document explains each point and turns it into concrete tasks.
How these functions work on the inside
Understanding two techniques helps to decide what is worth investing effort in.
Recovery Augmented Generation (RAG)
It is also known as a foundation. The model does not respond only from memory: Google’s main positioning systems extract relevant and up-to-date pages from the Search index, and the response is built by reviewing the specific information on those pages. That’s why the answer is more accurate and more current, and that’s why clickable links appear to the pages that support what he says.
The practical consequence is direct. If your page is not indexed and is not suitable to be displayed with a snippet, it does not enter the conversation.
Search branching
Based on the user’s query, the model launches several related queries at the same time to gather more results. Google gives this example: in front of “how to fix a lawn full of weeds”, the system can also search for “best lawn herbicides”, “eliminate weeds without chemicals” and “how to avoid weeds on the lawn”.
A single user query becomes a range of searches. Your page can appear by any of those branches, not just by the original query. Covering a topic well, with real depth, is the sensible way to take advantage of it. Creating a page for each possible variant is another thing, and Google advises against it (see section 4).
AEO and GEO according to Google
AEO stands for answer search engine optimization and GEO stands for generative search engine optimization. Both terms describe work focused on improving visibility in search experiences with AI.
Google’s position is that, from their perspective, optimizing for generative search is optimizing for the search experience, and therefore it’s still SEO. Anyone who hires third-party services on these issues should contrast the advice with the official guidelines to evaluate SEO recommendations.
This does not invalidate the visibility work on other AI platforms. Talk about what’s going on inside Google Search. If you optimize for other systems, which use their own criteria, this guide does not cover them.
Content best practices
Google claims that creating content that users consider unique, attractive and useful is what most influences a site’s presence in generative search. More than any other suggestion in the guide.
Offers a unique point of view
AI systems analyze many sources. A summary of what has already been published repeats what exists on other sites and gives no reason to quote you. A first-hand review, on the other hand, provides a perspective based on personal experience.
Ask yourself what you know about this topic that is not already in the first ten pages that cover it. If the answer is “nothing”, the content is not ready yet.
Avoid generic content
Google gives two examples that should be saved. “7 tips for first-time homebuyers” is generic content – common knowledge that anyone could write. “Why we forgo inspection and save money: A look inside the sewer pipe” is not generic, because it offers an expert perspective or own experience that goes beyond the conventional.
A quick test before posting: if a generative AI model could produce the text without anyone contributing anything, it’s generic.
Organize the content for humans
Write for the audience, with paragraphs, sections and headings that give a clear structure to navigate. There is no special recipe for AI. What helps a person to scroll through the page is what Google recommends.
Add quality images and videos
Generative AI features can include relevant images and videos, so your site has more places to appear in addition to links. Reinforce the text with quality visual material when it makes sense. If you already follow the SEO practices for images and for video, you are already optimizing for this.
Don’t go overboard with the variants
It can be tempting to create separate content for each query variation, including those that come out of the search branch. If you’re doing it primarily to manipulate the AI’s positioning or responses, it violates the large-scale content abuse spam policy. In addition, having many pages does not make a site better, and Google’s systems already understand relevance even without an exact match between query and content.
The basic criterion
Google proposes a question for any dubious decision: “Is this content that my visitors would find satisfactory?”. If the answer is yes, you are on the right track.
Technical structure
The way Google finds and processes your pages is still the basis of how AI systems access your data. All the usual technical SEO is still valid.
Eligibility requirements
- The page must be indexed and suitable to be displayed in the Search with a fragment.
- You must meet the technical requirements of the Search.
- The site must be included in the generative AI functions of Search in Search Console.
Complying with everything does not guarantee that Google will crawl, index or serve the content. There is no guarantee of indexing or publication.
Tracking
The content has to be trackable, because models use public, trackable content to learn patterns and give informed responses. On very large sites that are updated frequently, check out Google’s guide to crawl budget.
Semantic HTML
There is no need for perfect HTML, and Google can understand the web as it is. Still, semantic HTML helps screen readers and other users navigate the page, and it’s usually a good idea to use it when you can. The criterion is human-readable, not the purity of the code.
JavaScript
Google may process content in JavaScript as long as your access is not blocked. Even so, SEO on sites with JavaScript frameworks is usually more complex, so it is advisable to follow the usual SEO practices for JavaScript.
Experience on the website
The site should look good on all devices, have low latency and make it easier for the user to distinguish the main content from other elements.
Duplicate content
It generates a bad experience and can cause search engines to spend tracking resources on URLs that do not interest you. If you have time, reduce it.
Diagnosis
Search Console is the tool to detect and diagnose technical problems quickly.
Local business and e-commerce
Generative AI responses can include product sheets, product information, and local company data. Two tools help to appear there:
- Merchant Center, including its feeds, for products.
- Company profiles on Google, for local businesses.
Depending on the type of business, there are other experiences for merchants, such as the Commercial Agent, a conversational experience in the Search that helps customers chat with the brand.
If you have an online store or a business with a location, this is a concrete and measurable work front, unlike many speculative tactics.
Myths that Google says you can ignore
Google collected what is circulating on the internet and points out what does not need to be done in Google Search.
| Myth | What does Google say |
|---|---|
| Create LLMS files.txt or other labeled “special” | No need for machine-readable files, AI text files, tags or Markdown. Google Search completely ignores them. If you keep them for other services, they do not harm or benefit your positioning. |
| Breaking the content into small parts | It’s not necessary. The systems understand several topics on one page and show the relevant part. There is no ideal length. |
| Rewriting content only for AI systems | There is no need to write specifically for AI. They understand synonyms and the general sense of what the user is looking for, so you don’t need to cover every exact variant. |
| Search for inauthentic mentions | AI functions can show what is being said about products and services on the web, but classification systems focus on high-quality content and other systems block spam. Artificial mentions yield less than it seems. |
| Obsessing over structured data | Are not mandatory for the generative search and no special labeling is needed for schema.org . They are still a good idea as part of the general SEO, because they help to opt for enriched results. |
An important nuance. Structured data does not mean that it is useless. He says that they are not a requirement to appear in AI functions, and that their value comes from the enriched results. And about LLMS.txt is limited to Google Search: other services can use them.
Measurement
To see how your content is performing in Search’s generative AI features and in Discover, use the Search Console Generative AI Performance report. Here you see how users discover your content through these functions.
About third-party tools, Google asks for distrust with those that promise to improve positioning or claim to use “internal” Google metrics. None have access to their internal positioning or AI systems. You can use them if they help you in your workflow, but contrast their advice with the official guidelines.
Using AI to create content
Generative AI is used to investigate a topic and give structure to original content. The problem appears when it is used to generate many pages without adding value, because it may violate the spam policy of large-scale content abuse.
Google recommends reading two sections of the guidelines for quality reviewers: 4.6.5, on large-scale content abuse, and 4.6.6, on main content created effortlessly, not original or with little added value. These guidelines are not a guide to positioning. They are used by evaluators to measure the performance of systems, and their evaluations do not directly influence positioning.
9.1 Accuracy and revision
Generative models do not retrieve data, they predict a likely sequence of words from their training. That is why they can produce inaccuracies, known as hallucinations. You have to manually verify the data and review all the content before publishing it.
The review includes the metadata that may appear in results:
- Elements of the title.
- Meta descriptions.
- Structured data.
- Alternative texts of images.
The structured data must also comply with the general guidelines, the policies of each search function and be validated so that they can be used in enriched results.
Give context to the reader
If you generate content automatically, Google recommends explaining how it was created, in a way that makes sense to your audience. For example, indicating how the automation was used and adding metadata to the images.
Requirements in e-commerce
Google Merchant Center has its own policies on AI content:
- AI-generated images must include the DigitalSourceType value TrainedAlgorithmicMedia in their metadata.
- AI-generated product data such as title and description should be specified separately and labeled as AI-generated.
Experiences with agents
AI agents are autonomous systems that do tasks on behalf of people, such as reserving a table or comparing product specifications. Browser agents can access a site to collect the data they need: they analyze the visual rendering (for example with screenshots), inspect the DOM structure and interpret the accessibility tree.
Google poses it as something optional: if you are interested and have time, check out the available agentic experiences and its best practices guide for agent-compatible sites. He also mentions emerging protocols such as the Universal Commerce Protocol (UCP), which will allow Search agents to do more.
A practical reading: a site with semantic HTML, clear structure and good accessibility already has an advantage with these agents, because they interpret just those layers.
Action plan by priority
High priority
- Confirm in Search Console that the key pages are indexed and are suitable to be displayed with fragment.
- Verify that the site is included in the generative AI functions of the Search in Search Console.
- Audit the existing content and mark the generic pages, that is, the ones that repeat what everyone already says.
- Rewrite those pages with own experience: data, real cases, evidence, informed opinion.
Medium priority
- Review tracking, duplicate content, performance and mobile viewing.
- Improve headings and organization of long pages so that they can be scrolled through.
- Add your own and relevant images and videos, following SEO practices for images and video.
- If there is a local store or business, set up and maintain Merchant Center and the Company Profile.
- Manually review all content made with the help of AI, including title, meta descriptions, schema and alternative texts.
Low priority
- Keep structured data as part of the overall SEO, for the enriched results.
- Follow the evolution of experiences with agents and protocols such as UCP.
- If you keep llms.txt for other services, leave it, but don’t count it as SEO work for Google.
What to stop doing
- Produce pages for each query variant to “cover” ramifications.
- Cutting content into minimal blocks with AI in mind.
- Buy or manufacture mentions.
- Rewriting pages just to make them “sound” like AI.
Quick checklist
| Area | Control question | State |
|---|---|---|
| Indexing | Are the important pages indexed and with fragment? | |
| Consola de Búsqueda | Is the site included in the generative AI features? | |
| Contents | Does each page contribute something that is not already in the other results? | |
| Contents | Is there any first-hand experience, data or own criteria? | |
| Structure | Do headings and sections allow to navigate the page? | |
| Media | Are there own and relevant images and videos? | |
| Technical | Tracking, duplicates, mobile and speed under control? | |
| JavaScript | Is the content rendered and access is not blocked? | |
| Trade | Are Merchant Center and Company Profile up to date? | |
| AI in the content | ¿Todo se revisó a mano, metadatos incluidos? | |
| Tagged IA | Do AI-generated product images and data comply with Merchant Center policies? | |
| Measurement | Is the generative AI performance report being reviewed? |
How to explain it to a client
The question that is going to appear the most is whether it is necessary to hire something different for AI. The honest answer that emerges from the Google guide is that, for Google Search, the work is the same as always, done with more rigor: content that really contributes, a technically sound site and measurement in Search Console.
Three sentences that summarize the position:
- LaGoogle’s AI functions are powered by the same positioning and quality systems as traditional SEO.
- There is not a file or a trick that replaces a useful and own content.
- What we can measure is the performance in the generative AI report of Search Console.
If a provider promises results with “internal Google metrics,” you should ask them to prove it, because Google claims that no third-party tools have access to those systems.
Glossary
- AEO: optimization of search engines for answers.
- GEO: generative search engine optimization.
- RAG: generation augmented by retrieval; substantiation technique that brings relevant pages from the index to support the answer.
- Search branching: Related queries that the model launches in parallel from the original query.
- Views created with AI and AI mode: Generative AI functions of Google Search.
- UCP: Universal Commerce Protocol, designed so that search agents can do more tasks.
- Browser Agent: Autonomous system that accesses a site to complete tasks on behalf of a person.
Closure
The most uncomfortable part of the guide is also the most useful: there is no shortcut. Google insists that there is a lot of content that performs well in Search, including AI experiences, without any obvious SEO, and that you don’t have to comply with everything the document says to be successful. What is repeated in each section is the same: own, useful and verifiable content, on a clean technical basis.