Google's AI Search Guide: What to Do and What to Skip (llms.txt, Chunking and Special Markup Myths)
According to Google's May 2026 guide, AI Overviews and AI Mode rest on the core ranking systems; an indexed page eligible for a snippet is enough, and llms.txt, chunking, rewriting for AI and special schema aren't needed. We cover the guide's "you don't need to" list, its emphasis on non-commodity content, spam policies applied to AI answers, and why the guide applies to Google Search only.
Version note: This post is based on the guide Google published for its generative AI (Artificial Intelligence) features, "Optimizing your website for generative AI features on Google Search" (May 15, 2026, llms.txt note June 15, 2026) and the Search Central change log as of November 2026.
Guess First
Four proposals are on the table at the Bookstore team's "AI search" meeting: (1) add an llms.txt file to the site, (2) split every blog post into independent 150-word pieces, because "AI reads in chunks", (3) add a special schema.org markup "for AI" to every page, (4) produce artificial "mentions" of Bookstore on forums and social media. Which of these do you think works for Google's AI Overviews and AI Mode? Which one could do harm?
The Guide's Core Claim: This Is Still SEO
On May 15, 2026 Google published a separate optimization guide for its generative AI features. The first section asks whether SEO (Search Engine Optimization) is still relevant: "Is SEO still relevant for generative AI search?" The answer is yes. According to Google, its generative AI features rest on the core Search ranking and quality systems; they use RAG (Retrieval-Augmented Generation), that is, grounding, and query fan-out. The guide refers to AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization) as names for the effort to improve visibility in AI search and ties the conclusion together in one sentence: optimizing for generative AI search is optimizing for the search experience, and therefore still SEO.
The rest of the guide unpacks that claim: foundational SEO practices, things explicitly considered unnecessary, measurement, agentic experiences and a "what to focus on" list. This post reads the guide in order and makes each item concrete on Bookstore.
Eligibility: Index, Snippet and the Search Console Setting
For a page to appear as a supporting link in AI Overviews or AI Mode, the guide lists two conditions. First, the page must be indexed and eligible to be shown in Google Search with a snippet, that is, meet the technical requirements. Second, the site must be included in the "Search generative AI" features in Search Console. The second condition relates to the site-level control introduced in June 2026; it's an opt-out, and the site is included unless you turn it off. The usual warning applies here too: meeting these doesn't guarantee crawling, indexing or serving.
The practical consequence: the first answer to "why isn't Bookstore in the AI Overview?" is in the diagnosis tree of the SEO Fundamentals series. If the page isn't indexed, carries nosnippet, or the site has been excluded from these features in Search Console, it can't get into AI features no matter how good the content is. The nosnippet, max-snippet and data-nosnippet controls are in the third post of the series.
Google's "You Don't Need To" List
The most discussed section of the guide is titled "Mythbusting generative AI search: what you don't need to do". The four opening proposals are almost one-to-one on this list:
| Proposal | Google's guide | What it means for Bookstore |
|---|---|---|
| llms.txt, AI text files, Markdown versions | Google Search doesn't use them; not needed to appear | Not a job for Google; evaluate separately for other services |
| Splitting content into small chunks | Not needed; the systems can handle multiple topics on a page, there's no ideal page length | Structure posts for readers, don't fragment them for the model |
| Rewriting content just for AI | No special phrasing needed; synonyms and meaning are understood, don't chase every long-tail variation | Don't produce an "AI version" of the same post |
| Artificial "mentions" | Not as useful as they seem; spam systems filter out manipulation | Spend the mention campaign budget on content |
| Structured data for AI | Not required for generative AI search, there's no AI-specific schema.org markup | Use structured data for rich results |
The answer to the opening question: none of the four proposals is necessary for Google's AI features. The fourth is also risky: producing artificial mentions gets close to the link spam and scaled content abuse lines, and the guide itself says spam systems filter out that kind of manipulation.
The chunking item has a subtlety. In the "Chunking Strategies" post of the RAG series we covered splitting documents for your own retrieval system; that is a design decision in a pipeline you control. What Google says here is something else: you don't have to split your content for Google's own systems. Structuring posts with clear headings and paragraphs is still a readability recommendation that also appears in Google's guide.
The llms.txt Note: What Does "Completely Fine" Mean?
On June 15, 2026 a note was added to the guide. The change log's summary: Search doesn't need llms.txt files and they don't affect rankings, but sites may keep them for other systems. The note in the guide says the same thing: it's "completely fine" if you decide to create and maintain llms.txt files, but doing so neither helps nor harms your visibility in Google Search, because Google Search ignores these files.
The note is open to two opposite misreadings. The first is "Google killed llms.txt"; the note only speaks for Google Search. The second is "Google endorsed llms.txt"; the note only says it's harmless. mertozen.com already publishes /llms.txt and /llms-full.txt; we'll cover the spec of these files, who uses them and how to generate them in ASP.NET Core in the fifth post of the series, with honesty about "what we can measure and what we can't".
Non-commodity Content: What Google Actually Wants
The first and longest item on the guide's "what to do" side is about content: create valuable, non-commodity content for your audience. Commodity content is built on common knowledge and adds little unique perspective. Google's example is a list like "7 Tips for First-Time Homebuyers"; its non-commodity example is a piece coming from expertise or experience, about something someone actually lived through, along the lines of "why we waived the inspection and how we saved money".
Applied to Bookstore, the difference is clear. "7 books for learning async programming" is commodity: hundreds of sites publish the same list, and an AI answer engine can take that information from any of them. "We tried the backpressure example from Async in Practice in our own order service: what changed, where we got stuck" is non-commodity: a measured, dated experience that exists nowhere else. The content an engine has a reason to cite is the second one.
The guide's other foundational items are the same as what we discussed in the SEO Fundamentals series: keep content crawlable, follow JavaScript SEO basics, provide a good page experience (display across devices, low latency, clear separation of main content from other elements), reduce duplicate content, verify in Search Console. The guide also recommends high-quality, relevant images and video, which can create opportunities to appear beyond text links, and keeping Merchant Center and Business Profile information up to date for local businesses and e-commerce sites.
Spam, Manipulation and Third-Party GEO Promises
On May 15, 2026, the same day as the guide, Google stated explicitly in its documentation that the spam policies also apply to generative AI responses: attempts to manipulate generative AI responses in Google Search are spam too. The guide adds that producing many pages primarily to manipulate rankings or AI responses violates the scaled content abuse policy, and that the number of pages doesn't make a site higher quality.
This series doesn't teach tactics aimed at manipulating AI answers, such as putting hidden text or hidden instructions aimed at models on a page. They are AI-adapted versions of known violations in Google's spam policies such as hidden text and cloaking, and they put a site's reputation across all of search at risk. The indirect prompt injection topic in the "Authorization and Security in RAG" post of the RAG series explains why such content is also a security problem.
For third-party services, the guide points to the third-party SEO advice guidance published in June 2026 and draws the line in one sentence: no third-party tool has access to Google's internal ranking or AI systems. A service that claims an "AI Overview placement guarantee" or knowledge of Google's internal metrics therefore deserves skepticism.
The Guide's Scope: Google Search Only
This may be the most important warning in this post: every sentence of the guide is written for Google Search's AI features. The sentence "you don't need llms.txt" says Google Search doesn't use the file; it says nothing about what ChatGPT, Claude, Perplexity or a coding assistant does. Likewise, Google's measurement tools only see Google's surfaces.
Each of the other engines has to be read through its own documentation. OpenAI's, Anthropic's and Perplexity's documentation is detailed on bots and robots.txt behavior, but none has a guide as detailed as Google's about content format or ranking. Third-party "GEO playbooks" fill that gap; we'll compare them with the findings of independent research in the fourth post of this series.
Production Notes
Don't build a separate "AI pipeline" for Google. Producing an AI version, a chunked copy or special markup of the same content brings no benefit on the Google side; it brings maintenance load and duplicate risk.
Record the generative AI setting in Search Console as a decision. The setting is site-level and covers the generative AI features in AI Overviews, AI Mode and Discover together; someone turning it off without knowing means leaving all those surfaces.
Follow the guide's change history. The guide was published in May and the llms.txt note was added in June; documentation in this area changes fast. Checking the Search Central change log once a month is enough.
Common Pitfalls
Generalizing Google's "you don't need to" list to all AI engines. The list is for Google Search only.
Reading "not needed" as "forbidden". llms.txt or structured data do no harm on the Google side; they're just not a condition for appearing in AI features.
Confusing non-commodity with length. Originality doesn't come from word count; it comes from first-hand experience, measured data and a real point of view.
Looking for "tricks" specific to AI answers. In Google's explicit words, the spam policies apply to AI answers too; a shortcut is a risk for the whole site.
A Small Experiment
★☆☆ Label your site's last five posts as commodity or non-commodity and write a one-sentence justification for each.
★★☆ For a post you labeled commodity, write a plan to make it non-commodity by adding a real experience, measurement or decision from your team. Which data can only your team provide?
★★★ Extend this post's table for the other engines: find what OpenAI's, Anthropic's and Perplexity's documentation says about the same five topics. If the documentation says nothing, write "not in the docs" in the cell; don't fill it with guesses.
We've seen the Google side. In the next post we go down to the technical layer shared by all engines: separating AI bots into training, search and user-initiated fetches, and how to manage each with robots.txt.