Google AI Optimization Guide:
Do SEO, Don’t Game the Overviews

Published May 15, 2026

On May 15, 2026, Google published two documents that, read together, settle a two-year-old argument inside the search industry. The first is a brand-new Google AI Optimization Guide. The second is a refreshed Web Search Spam Policies page that, for the first time, names generative AI answers as a surface its enforcement rules cover.

The combined message is simple. AI Overviews and AI Mode are not a separate game with separate rules. The same ranking systems decide what appears there, and the same spam policies govern how you are allowed to try to appear there.

For the SEO industry, this closes a door. The “what to do” and the “what gets you penalized” now come straight from Google’s own documentation, and neither page mentions AEO, GEO, llms.txt, or any of the parallel vocabularies the vendor market has been selling since 2024.

What Google says to do

The AI Optimization Guide reads less like a new playbook and more like a restatement of SEO fundamentals, applied to AI-powered answers.

The advice is short:

• Write original, first-hand content. Google’s own example contrasts “Why We Waived the Inspection & Saved Money” with “7 Tips for First-Time Homebuyers.” The first has lived experience inside it. The second is a commodity. AI Overviews already summarize commodity advice on their own, so pages that just repeat that material have nothing extra to contribute to the answer.
• Structure your pages for human readers. Clear sections, headings, and a readable layout. Nothing engineered for a hypothetical AI parser. The same readability cues that help a human skim a page also help Google’s systems pull the right passage into an AI answer.
• Keep the site technically healthy. Crawlable, indexable, fast, with JavaScript rendered properly. Use Search Console to find what is broken. If Google cannot index a page in the first place, none of it can show up in an AI answer.
• Take care of your business surfaces. Merchant Center feeds and Google Business Profile carry product and local visibility into AI answers. Google also points to its new Business Agent for conversational commerce. For ecommerce and local businesses, these feeds are the structured data path Google does actually read.
• AI-assisted writing is allowed, as long as the result meets Search Essentials and the spam policies. What matters is the page itself: AI-assisted copy with real expertise behind it is fine; AI-generated filler is not.

None of that is new. That is the point.

What Google says to stop doing

The more interesting half of the guide is the list of popular “AI SEO” tactics Google describes as ineffective:

• llms.txt files.  Google does not use them. The file was a community proposal from 2024, never adopted by Google, and this guide settles the question
• Breaking content into “AI-readable” chunks. Not needed. Google’s systems can extract a relevant passage from a long article on their own, the same way they have always pulled featured snippets.
• Rewriting copy in an “AI-friendly” voice. Not needed. Models understand synonyms and natural phrasing — they do not reward writing that sounds like it was prepared for a machine.
• Chasing every long-tail variation of a query. Not needed. Semantic matching means one well-written page can answer many related queries without listing every variant in the copy.
• Paying for or seeding fake brand and product mentions across the web. Does not work. Google’s systems already weight authentic mentions higher than coordinated ones, and this category now sits inside spam policy.
• Special schema markup aimed at generative AI. Does not exist. No new schema.org type or attribute was created for AI Overviews, and the existing markup remains optional for appearing in AI answers.

Google says it plainly: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in generative AI search.”

This is the part the SEO vendor market has been quietly avoiding for the last twelve months. A large chunk of the GEO and AEO consulting market has been built around exactly the items on that list.

Manipulating AI Overviews is now officially spam

The companion document is shorter, but it carries more weight.

Google’s Web Search Spam Policies page now defines spam as including “attempting to manipulate generative AI responses in Google Search.” That single phrase moves AI Overviews and AI Mode inside the enforcement perimeter of Google’s existing spam framework — alongside cloaking, doorway abuse, hidden text, link spam, scraped content, sneaky redirects, thin affiliate pages, and the rest of the familiar list.

One existing category deserves singling out. Under scaled content abuse, Google now spells out that “using generative AI tools or other similar tools to generate many pages without adding value for users” counts as a violation. That is the back-end of half the AI content operations currently running on the web.

Enforcement looks the same as before: automated detection backed by manual actions, with offending sites ranked lower or removed from results. What changed is the surface area. The same rules apply to what shows up in AI answers, not just the ten blue links.

Read against the Optimization Guide, the practical consequence is clear. Techniques being marketed as GEO — fake citations, scaled AI page generation, third-party content slipped under a trusted host’s reputation, seeded brand mentions — are not just ineffective. Aimed at AI answers, they are now spam-policy violations.

What’s actually new in the Google AI optimization guide

The two documents do flag a few real shifts:

• Exact-match keywords matter less. Semantic understanding does more of the work.
• Page length has no ideal target. Write what the topic actually needs.
• Structured data still helps with rich results, but is not a requirement for AI surfaces.
• Google nods at “agentic experiences,” where AI agents interact with sites via the DOM and the accessibility tree, and points to its Universal Commerce Protocol for that future.

These are evolutions of SEO, not replacements for it.

AEO and GEO were always expanded SEO

If Google’s documents land like a confirmation, that is because someone said this out loud first.

Pedro Dias, writing in The Inference (The Whole Point Was the Mess), laid out the argument months before Google did. He put it bluntly:

“SEO-era controllability, repackaged under a new acronym. The same percentages, pillars, and pyramids.“

The mechanism point Dias makes is the one Google’s guide now confirms:

“There is no parser inside the model looking for <schema> tags… The model reads the words. That is the mechanism.“

Dias also flags something the GEO vendor market would rather not discuss. The original academic paper that introduced the term GEO found that classic SEO keyword tactics “may not translate to success in this new paradigm.” The acronym got borrowed. The findings did not.

Lily Ray published her own research that added the field data. Her recent analysis of 220+ websites scaling content with AI (It Works Until It Doesn’t) found that 54% of those sites lost at least 30% of their peak organic traffic, with 22% losing 75% or more. Her framing of the link between SEO risk and AI search risk is direct:

“What is dangerous for SEO can also be dangerous for AI search, largely because of RAG.“

Ray’s list of risky AI content templates includes comparison pages, glossary entries, “best X for Y” listicles, self-promotional rankings, and programmatic location pages. That is the same list now being produced at scale and pitched as GEO content. The mechanism Dias describes and the traffic losses Ray measures are two views of the same outcome.

Dias’ diagnosis of the moment is this:

“What you are watching is what happens when an industry has been organized for a quarter-century around the premise that you can pull a lever and watch the meter move, and the meter quietly disconnects from the lever.“

Google’s two new pages are the official version many have already been thinking. The mess was always the point. The shortcuts around it now sit inside spam policy.

What to do on Monday morning

A short list for anyone responsible for organic visibility:

• Audit your top pages. Replace recycled summary content with first-hand expertise where you can.
• Drop any roadmap items for llms.txt files, AI-only schema, or “chunking for AI.”
• Fix indexation, JavaScript rendering, and Core Web Vitals before anything else.
• Get Merchant Center and Google Business Profile current if either applies to you.
• Cancel any brand-mention seeding services pitched as AI visibility. That is inside spam policy now.
• If you have been using AI to mass-publish thin pages, pull them down. Scaled content abuse names this directly.
• Link your team to both Google pages, the Pedro Dias piece, and Lily Ray’s data analysis.

The Bottom Line

The work has not changed, but the labels did (as they do every few years, when so-called experts try to repackage one thing and sell it as a different service. Google has now said both halves out loud on the same day – the path to AI Overviews is SEO, and the shortcuts around it are spam.