Google's official guide to AI Overviews, last updated July 10, 2026, says what most AEO and GEO advice won't: its AI features run on the same ranking systems as regular Search, and 5 popular tactics, llms.txt included, aren't needed there.
The document is called "Optimizing your website for generative AI features on Google Search", and it is the closest thing to a rulebook Google has published for AI Overviews and AI Mode. Below we quote it directly, sort the advice into what matters and what doesn't, and check it against our own numbers: 1.91 million AI Overview and AI Mode impressions over mid-July to mid-October 2026, documented in our AI search traffic case study. Where we disagree with or go beyond the guide, we say so.
Sources: Google Search Central, AI optimization guide (updated July 10, 2026); Value Add VC Search Console, Generative AI performance report, mid-July to mid-October 2026.
Google AI optimization guide: the short version
Google's AI optimization guide says SEO is still the job. AI Overviews and AI Mode pull from the Search index, so a page must be indexed and snippet-eligible, and unique, first-hand content matters more than anything else. It tells site owners to skip llms.txt, chunking, AI-only rewrites, fake mentions and special schema.
The guide opens with a direct question, "Is SEO still relevant for generative AI search?", and answers "In short, yes!" because the features "are rooted in our core Search ranking and quality systems." On the labels themselves, it says that "from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO." Our longer take on how the three labels relate is in SEO vs AEO vs GEO.
How do AI Overviews choose which sources to cite?
Google names two techniques. One is retrieval-augmented generation, also called grounding, which relies "on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index." The model then reviews those pages to write the answer and shows "prominent, clickable links" to the pages that support it.
The other is query fan-out: "a set of concurrent, related queries generated by the model to request more information." Google's example is a user asking how to fix a weedy lawn, which fans out into searches like "best herbicides for lawns" and "how to prevent weeds in lawn." Google's separate AI features page adds that AI Mode and AI Overviews "may use different models and techniques," so the links each shows will vary, and that AI Overviews "often don't trigger" when they would not add to classic results.
The practical consequence: a page gets cited in an AI answer by ranking for one of the many sub-queries behind it, not only for the question the user typed. In our data, that shows up as machine-shaped queries. Our blog drew 301K impressions from them at a 0.01% CTR in the 28 days to September 21, 2026, against 360K human-shaped impressions at 0.95%, per the results file of our open-source playbook. We cannot prove those are Google fan-out queries specifically; they read like prompts, which fits.
What does Google say actually matters?
Google ranks content quality first: unique, valuable content "will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide." It draws a line between commodity and non-commodity content. Its commodity example is "7 Tips for First-Time Homebuyers"; its non-commodity example is "Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line." It tells writers not to recycle what "could easily be produced by a generative AI model."
Technical access comes second. To be eligible, "a page must be indexed and eligible to be shown in Google Search with a snippet," and the site must be included in Search generative AI features in Search Console. Google lists crawlability, JavaScript that isn't blocked, a good page experience across devices, lower latency and less duplicate content. It says semantic HTML is "generally a good idea" but not required, since "the web in general is not valid HTML."
Third, media and business data: high-quality images and video can appear in AI answers, and Merchant Center feeds and Google Business Profiles feed product and local information into them. Google also flags a trap: creating separate pages for every fan-out variation, done "primarily to manipulate rankings or generative AI responses," violates its scaled content abuse spam policy.
What can you ignore for AI Overviews?
Google's "Mythbusting generative AI search" section lists 5 things you can ignore for Google Search, and the closing section adds a warning about third-party tools. Here they are against the claim each one answers.
| Common AEO/GEO claim | What Google's guide says | Verdict for Google |
|---|---|---|
| Add an llms.txt file or AI markup | "Google Search itself doesn't use them"; it will "neither harm nor help" | Ignore |
| Chunk pages into small blocks | "There's no requirement to break your content into tiny pieces" | Ignore |
| Rewrite content for AI and long-tail variants | "You don't need to write in a specific way just for generative AI search" | Ignore |
| Buy or seed mentions across the web | Seeking inauthentic mentions "isn't as helpful as it might seem"; spam systems apply | Ignore |
| Special schema gets you into AI answers | "There's no special schema.org markup you need to add" | Keep for rich results only |
| A page per fan-out query | Can violate the scaled content abuse spam policy | Avoid |
| Tools that see Google's internal AI metrics | "No third-party tool has access to our internal ranking or AI systems" | Be wary |
Source: Google Search Central, "Optimizing your website for generative AI features on Google Search," last updated July 10, 2026. Quotes are verbatim; the verdict column is Value Add VC's summary.
How we counted: every tactic the guide's mythbusting and measurement sections address; 1 document; 7 claims; as of October 16, 2026.
Is llms.txt useless, then?
For Google Search, yes, by Google's own account. Outside Google, the picture differs. The guide says it is "completely fine" to maintain llms.txt "for other services or systems that use these files," and some AI companies publish one themselves: Perplexity's crawler documentation opens by pointing AI agents to its docs index at /llms.txt. Value Add VC keeps one too, and we credit none of our Google AI impressions to it. Our llms.txt explainer covers when it is worth keeping one.
Three common misreadings of the guide
"Ignore structured data." The guide says not to overfocus on it for AI answers, then says "it's a good idea to continue using it" because it helps pages qualify for rich results. Removing working schema to save time would be a mistake.
"Don't chunk, so structure doesn't matter." The same guide tells writers to organize pages "by paragraphs and sections, along with headings that provide a clear structure." The point is that you need not break content into tiny AI-sized fragments, not that a wall of text is fine. Readable structure is for people, and Google says that is enough.
"It's still SEO, so nothing changed." Ranking inputs may be the same, but the surface is not. Fan-out means one user question can trigger many searches, and an AI answer can satisfy a reader without a click. Measuring has changed even if optimizing hasn't, which is why Google built a separate Search Console report and why our AI referrals and AI impressions moved in opposite directions this autumn.
What does Google say about AI agents?
The guide ends with a short section on agentic experiences: AI agents that act for people, "such as booking a reservation or comparing product specifications." Google says browser agents may read a site by "analyzing visual renderings (like screenshots), inspecting the DOM structure, and interpreting the accessibility tree," and names the Universal Commerce Protocol as one emerging standard that will let Search agents do more. It frames this as optional work "if you have extra time." Our read: the accessibility tree is the useful part, because accessible markup serves screen readers, agents and Google at once.
Does following the guide work? Our data
Our results are consistent with the guide, with one caveat about cause and effect. Value Add VC's weekday AI Overview and AI Mode impressions rose from about 10,000 a day in July 2026 to about 40,000 a day in October, with no AI-specific rewrites or chunking. Google Search overall sent 60.9K clicks from 13.3 million impressions in the same three months.
The pages AI sends readers to on our site are non-commodity by Google's definition: sourced money and data explainers such as our AI spending tracker and Anthropic profitability analysis, each built on figures we pulled and checked rather than restated. The one technical rule we learned the hard way is the indexing requirement: a site-wide noindex header live for about 28 hours on September 16-17, 2026 cut our daily impressions about 55% at the low point.
What the guide doesn't cover
The guide is about Google only. ChatGPT, Perplexity and Claude run their own crawlers and their own rules. OpenAI's crawler documentation says sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers," and Perplexity recommends allowing PerplexityBot to appear in its results. For those engines, robots.txt and firewall settings are a visibility decision in their own right, which is where AEO and GEO advice has a real job to do. We cover the crawlers and how to log them in our AEO and GEO tracking guide.
Where we could be wrong
Google has reasons to say "just do SEO": it keeps site owners inside its own measurement tools and discourages tactics that game its answers. The guide also offers no click data, and Search Console's AI report counts impressions only, so nobody outside Google can confirm how often an AI citation turns into a visit. Our own AI-assistant referrals fell 28% in early October while impressions grew. Taking the guide at its word on what gets cited is reasonable; taking it as proof that citations pay is not.
A checklist built from the guide
- Confirm key pages are indexed and not blocked from snippets (no nosnippet or noindex by accident).
- Check that your site is included in Search generative AI features in Search Console.
- Audit for commodity pages: anything a model could write from common knowledge.
- Add first-hand data, tests or experience to the pages you most want cited.
- Keep structured data for rich results, without expecting it to win AI citations.
- Skip llms.txt for Google; keep it only if another engine or agent you care about reads it.
- Watch the Generative AI performance report monthly and log what changed on your site.
Items 1, 2 and 7 come straight from Google's guide; the rest is our application of it. Our AI SEO playbook has the scripts we use for the indexing and query checks.
Latest from the Pulse
Get VC data most people never see
— free to subscribe
Trace's notes on venture, AI and startups, a few times a week. Join 5,000+ subscribers. No spam.
No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗