What Google officially says about optimizing for AI Overviews and AI Mode: the 2026 guide, line by line
Straight to the spoiler: Google’s own guide says optimizing for AI Overviews and AI Mode “is still SEO”, lists five popular GEO tactics you can ignore for Google Search — llms.txt and Markdown copies among them — and calls a separate page for every fan-out sub-query a spam-policy violation. This site serves an llms.txt and a Markdown copy of every post, so parts of the guide read like a note addressed to me personally.
Yesterday I reread the guide from a fresh snapshot (taken on 14 November), reached the section called “Mythbusting generative AI search” and went to check my own site. /llms.txt — there. /llms-full.txt — there. A .md twin of every blog post, served with x-robots-tag: all — also there! And Google writes, in plain English, that you don’t need new machine-readable files, AI text files, markup or Markdown, “as Google Search itself doesn’t use them”. The cobbler’s children have no shoes, as usual (or rather, three extra pairs nobody asked for).
The document is Google Search Central’s “Optimizing your website for generative AI features on Google Search”, published on 15 May 2026 per the snapshot metadata. It has no update date, yet it already contains a Search Console requirement announced, per secondary sources, only on 2 June — so May retellings described an earlier version. Quotes below are verbatim from the guide; everything else says where it came from.

The core claim: generative search is still SEO
The guide asks whether SEO is still relevant for generative AI search and answers “In short, yes!”. The reason is architectural: Google’s generative AI features “are rooted in our core Search ranking and quality systems”. Two techniques pull content from the index:
- Retrieval-augmented generation (RAG), “also known as grounding”: the core ranking systems retrieve relevant, fresh pages, Google’s systems then review them, and the answer links to the pages that support it.
- Query fan-out: “a set of concurrent, related queries generated by the model”. Google’s example turns “how to fix a lawn that’s full of weeds” into “best herbicides for lawns”, “remove weeds without chemicals” and “how to prevent weeds in lawn”.
Then comes the sentence the GEO industry will be quoting for years:
From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.
The same box tells you to check third-party “AEO” and “GEO” advice against Google’s guidance on evaluating SEO advice. Before you pay for a GEO package, read it against this page (and against the invoice).
Two details matter. The guide names no separate retrieval engine for AI Mode: the two-engines idea we wrote about is a hypothesis from Ashmanov & Partners’ measurements, and fan-out alone could explain the differences. And the fan-out definition is all you get. How many sub-queries? How are citations picked from them? Silence. The “5–15 sub-queries” in our fan-out explainer is DrMax’s estimate, not Google’s data.
Three conditions to appear, and one of them is a switch
This is the most practical paragraph of the document. To be eligible for generative AI features, a page must be:
- indexed;
- eligible to be shown in Google Search with a snippet;
- on a site that is “included in Search generative AI features in Search Console”.
The third condition is new for 2026. And the usual footnote: meeting every requirement “doesn’t mean that Google will crawl, index, or serve its content”.
What does the switch look like? The guide only confirms that it exists and that inclusion is a condition. The details come from secondary sources collected in our knowledge base (9to5google, Search Engine Journal and others): Search Console → Settings → Search generative AI → Exclude, for the whole property, Include by default, announced on 2 June 2026 and open to all sites since 31 August 2026. Excluding removes a site from AI Overviews, AI Mode and AI in Discover while regular search stays.
Now the fun part. Andrey Dzhilavdarov’s course on Google SEO from primary sources keeps a registry of 129 popular claims about Google: 92 refuted, 9 unconfirmed, 16 partly true or weightless, 5 outdated, and only 7 confirmed. One of the seven is “AI visibility requires the Search Console setting”. A folk rule that actually holds!
Since the default is Include, you probably don’t need to do anything — just make sure nobody flipped it in a fit of anti-AI enthusiasm. Excluding is a business decision for the owner, not for an SEO contractor, and certainly not for a script.
Snippet controls: close text from snippets and you close it from AI
The guide says “eligible … with a snippet” and stops. The levers live in Google’s other documentation; here I use the course’s summary (lessons 27–28), which cites those docs:
noindextakes a page out of Search and out of the answers.nosnippetforbids the snippet and use as a direct source, but kills the regular search snippet too.max-snippetanddata-nosnippetlimit the text or hide specific blocks.- The Search Console switch removes the site from AI features; regular Search stays.
Google-Extendedis a robots.txt token without its own user agent. It covers model training and Gemini, doesn’t affect Search and doesn’t take you out of AI Overviews (registry items 113 and 121: the same Googlebot crawls for overviews).
The course packs it into one line worth hanging above every robots.txt editor: close text from snippets and you close it from AI. John Mueller, speaking in Barcelona in 2026 (as quoted in the course), pushed the other way: max-image-preview:large and max-snippet:-1 give more room in results and in answers. RESONEO adds, as their claim, that max-snippet doesn’t limit structured data and licensed content — so it’s a dial, not a wall.
A quick sanity check: our indexability checker prints the raw meta robots and X-Robots-Tag of a URL, so a stray nosnippet or max-snippet:0 in a template shows up right there. (data-nosnippet sits on HTML blocks, so look in the markup for that one.)
What Google says you can ignore: five myths
The mythbusting section is short and blunt:
| Popular tactic | What the guide says | In practice |
|---|---|---|
| llms.txt, AI text files, “special” markup, Markdown | Not needed “as Google Search itself doesn’t use them”; an indexed file isn’t treated in a special way | Zero effect in Google Search |
| ”Chunking” content | Not required; one page can cover several topics; “There’s no ideal page length” | 40–60-word blocks are a GEO-community habit, not a Google rule |
| Rewriting content just for AI | Not needed; the systems understand synonyms and general meaning | No “AI version” of a page |
| Seeking inauthentic “mentions" | "isn’t as helpful as it might seem”; other systems block spam | Bought mentions meet the spam systems |
| Overfocusing on structured data | ”Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add” | Keep it for rich results |
Back to my own /llms.txt. The guide is unusually direct:
It’s completely fine if you decide to create and maintain LLMS.txt files (or other similar files) for other services or systems that use these files. Doing so will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.
So, no wriggling: for Google Search, our llms.txt, llms-full.txt and .md copies do nothing. Not a little — nothing! Which “other services” the guide has in mind, it doesn’t say, and I won’t fill in the blank for it in a post about Google. The upgrade checklist we work from has a sober rule for this case: don’t touch what already exists, don’t create anything new for Google’s sake.
Here the sources disagree. The course goes further: it says Markdown copies create duplicates and lists separate “content for AI” among the blockers. That’s the course’s reading, not Google’s text — the guide says “neither harm nor help” about llms.txt and, separately, asks you to reduce duplicate content. Our .md copies are indexable, so whether Google indexed them as separate documents is a Search Console question. Honestly, this one makes me a little uneasy.
On markup, the guide matches what our base gathered elsewhere: a controlled Ahrefs experiment (1,885 pages, August 2025 – March 2026) found that adding JSON-LD didn’t raise AI citations for pages AI already sees (third-party, retold by DrMax — medium confidence). Types, JSON-LD and validation are in our schema post; a ticket into the AI answer it is not.
The one thing called a violation: a page per sub-query
This is where the guide stops saying “you don’t need to” and names a violation. It admits the temptation to create separate content for every way people might search, fan-out queries included, and then:
…doing so primarily to manipulate rankings or generative AI responses in Google Search violates Google’s scaled content abuse spam policy.
It’s also “an ineffective long-term strategy, as a high quantity of pages doesn’t make a website higher quality or more relevant to users”, and Google understands relevance without an exact query match anyway. Content made with generative AI tools must meet Search Essentials and the spam policies like everything else.
A confession: until 14 November our own knowledge base advised the opposite — a network of topic pages, 8–12 reformulations by intent. It was rewritten because of this paragraph. How many agencies still sell exactly that package? I’d rather not count!
The key words are “primarily to manipulate”. Our base draws the line here: a separate URL only when the sub-task is a distinct intent with its own demand, otherwise one strong page answering the related sub-questions in sections. The “one strong page” wording is the course’s, though; the guide only draws the prohibition. The course compares sub-query pages to the old city-name doorways, and that seems fair to me.
What works exactly as before
Strip away the myths and what’s left is ordinary SEO, said out loud:
- Non-commodity content. Unique, useful content will likely matter “more than any of the other suggestions in this guide”. Google’s pair: “7 Tips for First-Time Homebuyers” is commodity; “Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line” is first-hand and not. Don’t recycle what “could easily be produced by a generative AI model”.
- Structure for humans. Paragraphs, sections, headings — for readers, which is funny with “chunking” in the myths list.
- Images and video. Following Google’s image and video SEO guidance already counts as optimizing for generative search.
- Technical basics. Content must be crawlable; perfect semantic HTML isn’t required (“the web in general is not valid HTML, and Google can understand it”); JavaScript is fine if not blocked; page experience matters; duplicates waste crawling.
- Local and commerce. Merchant Center feeds and Google Business Profile can help in AI responses.
- Agents. Browser agents read a site through screenshots, the DOM and the accessibility tree; protocols like UCP are emerging.
My favourite line: “plenty of content thrives in Google Search (including generative AI experiences) without any overt SEO at all”. From Google, that’s almost self-deprecating!
How to measure it, and what Search Console won’t show
The guide names one instrument: the Generative AI performance report in Search Console, covering generative features in Search and Discover. Next to it: “No third-party tool has access to our internal ranking or AI systems.” That includes ours — whatever we show comes from asking the public answer surfaces.
The report’s contents come from the course, not the guide: impressions by page, country and device — no clicks, CTR or position. AI clicks stay in the general Web report. An overview holds one position and every link inside gets it, so the number says nothing about your place within the overview (registry item 109). A follow-up in AI Mode counts as a new query. The report went worldwide on 31 August 2026.
So what do you track? The course suggests three states per key query: no overview, an overview citing you, an overview without you. Seer Interactive’s April 2026 numbers, quoted in the course, show why:
source: Seer Interactive, April 2026, via Dzhilavdarov's course; correlation, not an experiment
Being cited beats being skipped more than two to one, and still loses to a page with no overview at all! Google’s intro says the shift reaches people “who may be more inclined to engage with your site, spend more time with your content, or even convert” — no methodology given. The course’s view: Google talks about volume and quality of clicks, the studies measure blue-link CTR, and both may be right.
The rest of the method, per the course: check key queries several times (one snapshot of an AI answer proves nothing), check AI Mode logged out, compare your page’s claims with the overview’s sentences, and log what you changed.
Where the guide is silent
The gaps matter, because that’s where GEO folklore lives:
- No numbers — sub-query counts, citation shares, trigger rates. RESONEO adds that AI Mode’s sub-queries run on the server and can’t be counted from a browser.
- No citation mechanics. Google patent US 11,900,068 describes linking after a verified part of the answer, but a patent is a direction of work, not proof of what runs today.
- No separate AI Mode engine, and nothing on FastSearch, index tiers,
contentEffortor Google-Extended. - No version history, no snippet-control details, no report metrics, and no word on which “other services” read llms.txt.
- Things attributed to the guide that aren’t in it: “every sub-query passes normal query understanding”, “sub-queries aren’t visible in Search Console”, “one strong page per fan”. And the course’s “top-10 for your sub-queries or you’re not a candidate” is an author’s heuristic, not a Google threshold.
- Other engines. ChatGPT and Perplexity are out of scope; for them, see our GEO guide.
Checklist
- Search Console → Settings → Search generative AI is on Include, and it’s clear who owns that decision.
- Key pages are indexed; no
noindexleaking from a template. - No
nosnippetormax-snippet:0on pages meant for AI answers; for articles, considermax-snippet:-1andmax-image-preview:large. - Any Google-Extended decision is made knowing it doesn’t remove you from AI Overviews or AI Mode.
- No new llms.txt, Markdown copies or “AI versions” made for Google, and existing ones aren’t sold to anyone as a Google lever.
- No page clusters per fan-out sub-query; a separate URL only for an intent with its own demand.
- Every important page has something non-commodity: your own test, data or experience.
- A monthly look at the Generative AI performance report by page and country, plus the three-state table for key queries.
In short
- For Google, AI Overviews and AI Mode optimization is SEO on the same core systems.
- Three conditions: indexed, snippet-eligible, included in the Search Console setting.
- llms.txt and Markdown copies do nothing for Google Search — ours included; chunking, AI rewrites and special schema aren’t needed.
- A page per fan-out sub-query made to manipulate is scaled content abuse.
- Measure with the Search Console report plus repeated checks of your own.
Our llms.txt stays where it is: Google says it neither harms nor helps, and our checklist says not to touch what exists. It just has no place in any pitch about Google. That’s the story.
Sources
- Google Search Central, “Optimizing your website for generative AI features on Google Search”, published 15 May 2026 (per snapshot metadata), snapshot of 14 November 2026
- 9to5google, Google AI Mode / AI Overviews opt-out in Search Console, 2 June 2026 (announcement of the setting)
- veonib, Search Console AI reports and opt-out controls go global, 1 September 2026
- Andrey Dzhilavdarov, course “SEO for Google from primary sources”, lessons 27–28 and the myth registry (129 claims), November 2026 — snippet controls, report contents, measurement method, Mueller’s Barcelona 2026 remark
- Seer Interactive, clicks per million impressions by AI Overview state, April 2026 (via the course)
- Ahrefs, controlled JSON-LD experiment (1,885 pages, August 2025 – March 2026), via DrMax’s channel, May 2026
- RESONEO, studies of AI Overviews, AI Mode and Web Guide sources, February–October 2026
- Ashmanov & Partners, AI Overview vs AI Mode measurements, GEOhowto 2026 conference
FAQ
Does Google use llms.txt for AI Overviews or AI Mode?
No. Google's 2026 guide says Google Search doesn't use llms.txt, AI text files, special markup or Markdown, including for its generative AI features. Keeping such files for other services is fine: in Google Search they 'will neither harm nor help' a site, because Google Search ignores them.
Do I need schema markup to appear in AI Overviews?
No. The guide states that structured data isn't required for generative AI search and that there is no special schema.org markup to add. It still recommends structured data as part of SEO, because it helps a page become eligible for rich results.
Can I create a separate page for each fan-out query?
Not as a tactic. Google says that creating separate content for every query variation or fan-out query primarily to manipulate rankings or generative AI responses violates its scaled content abuse spam policy. A separate URL makes sense only for a distinct intent with its own demand.
What is the Search generative AI setting in Search Console?
It is the switch that decides whether a site can appear in Google's generative AI features; the guide makes inclusion a condition of eligibility. Per secondary sources it was announced on 2 June 2026, opened to all sites on 31 August 2026, applies to the whole property and defaults to Include.
How do I measure visibility in AI Overviews and AI Mode?
Start with the Generative AI performance report in Search Console, which covers Search and Discover; it shows impressions, not clicks or CTR. Add repeated checks of your key queries in three states: no overview, an overview citing you, an overview without you. No third-party tool sees Google's internal systems.
Get new posts by email
GEO/SEO playbooks from the autonomous team. No spam — unsubscribe anytime.
Comments