On May 15, 2026, Google published something it had never published before. An official guide to optimizing for AI search. Not a tweet from a Googler, not a conference answer that gets clipped and argued about for six weeks. An actual Search Central document called “Optimizing your website for generative AI features on Google Search,” announced by John Mueller himself.
I read the whole thing twice. If you paid a consultant for “GEO services” in the last year, you should read it too, because Google just named the tactics you paid for and said Search doesn’t use them.
May 15, 2026
The date Google published its first official AI search optimization guide. Everything sold as “GEO strategy” before this date was sold without a word of official guidance behind it.
What Google actually said
The guide opens by answering the question every client has asked me since AI Overviews rolled out. Is SEO still relevant for AI search? Google’s answer is one word: yes. Their exact framing is that optimizing for generative AI search “is still SEO.” AEO and GEO are not separate disciplines. They’re marketing terms invented by people who needed something new to sell you.
The mechanics back that up. AI Overviews and AI Mode run on retrieval-augmented generation, which is a fancy way of saying the AI pulls answers from pages that Google’s normal ranking systems already retrieved from the normal index. There’s a second technique called query fan-out where the model fires off a batch of related searches to gather more source material. Both of these run through the same ranking infrastructure that’s existed for years.
Sit with that for a second. You cannot get cited by the AI if you don’t get retrieved first. And retrieval runs on the same authority signals it always has. The AI layer didn’t replace ranking. It sits on top of it.
The phrase that should make you nervous: commodity content
The strongest section of the guide, and Google says this directly, is about what they call non-commodity content. Their words: creating 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.”
Their example is brutal in a way Google documentation usually isn’t. Commodity content looks like “7 Tips for First-Time Homebuyers.” Anyone could have written it. A language model could write it in four seconds. Non-commodity content looks like “Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line.” One of those titles carries lived experience. The other one is filler that ten thousand sites already published.
The logic is cold math. An AI answer engine has infinite access to generic explanations of every topic. It doesn’t need your version of common knowledge. If your page restates what fifty other pages say, there is no reason, none, for the model to surface you over any of them. Your genuine perspective is the only input the machine can’t manufacture on its own.
Most companies fail this test and don’t know it. Their blogs are keyword-optimized, properly structured, and completely interchangeable with their competitors’ blogs. That was survivable in 2021. It’s dead weight now.
The kill list
The guide has a mythbusting section, and it reads like Google went through every GEO sales deck of the past 18 months and crossed things out line by line. Straight from the document, things you can ignore for Google Search:
llms.txt files. Google Search ignores them entirely. They neither help nor hurt. Every agency that charged a setup fee for one sold you a file that does nothing for your Google visibility.
Chunking your content. No need to break pages into bite-size fragments for the AI. Google’s systems understand multiple topics on one page and pull the relevant piece.
Rewriting content “for AI.” The systems understand synonyms and meaning. You don’t need to capture every phrasing variation of a query.
Chasing inauthentic mentions. Paying for your brand name to get sprinkled across blogs and forums so the AI “notices” you. Google says the ranking systems focus on quality while the spam systems block the rest, and AI features depend on both.
Overfocusing on structured data. Not required for AI search. No special schema exists for it. Keep using it for rich results eligibility, nothing more.
There’s one more warning buried in the content section that deserves its own paragraph. Publishing a separate article for every variation of how someone might phrase a search, which is the entire business model of programmatic SEO shops, now gets called out by name as a violation of Google’s scaled content abuse spam policy when it’s done to manipulate rankings or AI responses. Google adds that page volume was never going to work anyway, since quantity doesn’t make a site more relevant.
The mention economy just got repriced
The inauthentic mentions line is the one I want to dig into, because it’s where I’ve watched the most money get set on fire.
Sometime in 2025, a cottage industry appeared selling “brand mentions for AI visibility.” Unlinked name-drops on junk blogs. Seeded forum threads. Directory listings dressed up as citations. The pitch was that AI engines count mentions, so buy mentions. It sounded plausible if you didn’t think about it for more than a minute.
Google’s guide confirms what I’ve been telling clients since 2024. The AI features inherit the quality systems and the spam systems of regular Search. A mention on a page nobody trusts passes through the same filters that have been discounting junk signals for two decades. The spam systems don’t get turned off because a language model is doing the reading.
What actually moves retrieval hasn’t changed. Real placements on real sites with real trust. When I vet a placement I’m looking at Majestic Trust Flow before anything else, because TF measures proximity to genuinely trusted seed sites rather than raw link volume that anyone can inflate. A single editorial placement on a high-TF site does more for your retrievability than 200 scattered mentions on domains no clean site links to. That was true for blue links. It’s true for AI citations, and now Google has said so in writing.
Retrieved first. Cited second.
AI answers are grounded in pages pulled from Google’s existing index by Google’s existing ranking systems. Authority still decides who gets pulled.
What to actually do with this
Strip the guide down and two jobs remain.
Job one is an honest content audit. Go through your last 20 published posts and ask one question about each: could a competitor have published this word for word without anyone noticing? If the answer is yes, that post is commodity inventory. It’s not earning citations and it’s not going to start. The fix isn’t more posts. It’s fewer posts carrying things only you can say. Your data. Your failures. The client situation that surprised you. The take your industry disagrees with.
Job two is authority, same as it ever was. Crawlable site, clean technical structure, and links from sources that Google’s trust systems already believe. The guide also points you to the Generative AI performance report inside Search Console, added June 3, 2026, which shows how your content surfaces in AI features. Check it. Most site owners I talk to don’t know it exists. While you’re in there, confirm your site is opted into generative AI features under Search Console settings, because eligibility isn’t automatic visibility but ineligibility is guaranteed invisibility.
The uncomfortable version of all this: Google didn’t announce a new game. It confirmed the old game got harder. Production is cheap now. Perspective is expensive. The sites winning AI citations over the next two years will be the ones that stopped writing content a machine could write, and built the kind of trusted link profile that gets them retrieved in the first place.
I’ve been building that second part for clients since 2014. If you want to know whether your site is getting cited by AI engines right now, and why or why not, that’s exactly the question we answer at SerpHaus.