In May 2026, Google published official documentation on how websites get into AI Overviews and AI Mode. The short version: there are no special requirements, no secret markup, and no new discipline. If a vendor is selling you “GEO” or “AEO” as a separate line item, Google itself just told you to be skeptical.

I want to walk through what the document actually says, where most local business sites fail before any of it matters, and 5 checks you can run on your own site this afternoon.

What Google actually published

The document is called “Optimizing your website for generative AI features on Google Search.” It sits in Google’s official Search Central documentation, the same place they document how regular search works. It is not a blog post or a conference remark. It’s the reference.

Three statements in it matter more than the rest.

First: the AI features run on the same ranking systems as regular search. When Google generates an AI Overview, it retrieves pages from the normal search index and builds the answer from them. There is no second index and no second set of rules.

Second: the entry condition is the same as it has always been. Your page must be indexed, and it must be eligible to show a snippet. That’s it. No dedicated schema for AI, no special file, no markup that whispers to the model.

Third: Google addresses the “AEO” and “GEO” industry directly. Their position is that optimizing for generative AI search is optimizing for search (still SEO), and they point readers to their own guidance on evaluating third-party SEO advice before paying for any of it. When the platform that runs the results tells you to double-check the people selling access to those results, that’s worth taking at face value.

The part most sites fail before any of this matters

The entry condition sounds trivial. Indexed and snippet-eligible: of course your site is indexed, right?

In my experience, this is where local business sites actually lose, and for unglamorous reasons:

A WordPress install that takes 6 seconds to load on a phone, so Google crawls it reluctantly and ranks it accordingly.
The same site living at both yourbusiness.com and www.yourbusiness.com with no redirect between them, so every page competes against its own duplicate.
A staging site or an old subdomain still indexed, splitting authority three ways.
Structured data that was installed by a plugin in 2021 and describes pages that no longer exist.

None of this is exciting work. All of it decides whether you’re in the pool of pages an AI Overview can draw from. A vendor who opens the conversation with “AI visibility strategy” and not with your Search Console index report has the order backwards.

One concept worth understanding: query fan-out

The document describes something called query fan-out, and it’s the one genuinely new mechanic worth knowing. When someone asks Google’s AI a question, the system doesn’t run one search. It runs several related searches at once and assembles the answer from all of them.

Someone asks “how do I fix a lawn full of weeds,” and the system also quietly searches for herbicide options, chemical-free removal, and prevention.

For a local business, the implication is direct: covering your topic in one page is no longer enough, because the system is pulling from the pages that answer the neighboring questions too. A plumber with one page about water heater replacement competes worse than a plumber who also answered what replacement costs depend on, when repair beats replacement, and what permit rules apply in their county. Not 40 thin pages, but a handful of real ones that cover the territory.

The commodity content problem

The strongest section of Google’s guide is about content, and it draws a line I’ve been drawing with clients for 2 years.

Commodity content is the article any business in your industry could publish. Seven tips for this, five signs of that. Google’s own example contrasts a generic first-time homebuyer listicle with a first-person account of one specific transaction and the reasoning behind one specific decision. The first adds nothing to an AI-generated answer, because the model already knows the 7 tips. Everyone published them. The second is retrievable precisely because it exists nowhere else.

Here’s what most owners miss: this line favors you. A national content site can produce 400 commodity articles a month. It cannot produce the story of the water heater you replaced on Marengo Avenue last Tuesday, what the inspection actually found, and why you talked the customer out of the bigger unit. You can, and nobody can copy it.

The catch is extraction. That knowledge lives in your head and comes out in conversation, not in a writing session you’ll never schedule. In our own work, the solution was structural: a short weekly interview call, recorded, transcribed, and turned into the week’s content. The owner talks for 20 minutes. The system does the rest. However you solve it, solve it structurally: willpower-based publishing dies by March.

What I’ve watched break

Two failures from my own work, because a piece like this is dishonest without them.

The doorway page shortcut. Years ago the winning move in local search was generating a page for every service-times-city combination. Plumber in Pasadena, plumber in Arcadia, plumber in Monrovia: same page, swapped city name, 200 pages overnight. It worked until it catastrophically didn’t, and I’ve cleaned up the aftermath for a business that got caught in the correction. Under the fan-out model this approach is deader than ever: the system is looking for pages that add information, and 200 copies of one page add nothing 199 times. If a vendor proposes this in 2026, end the meeting.

Content that doesn’t ship. This year I watched a month of finished, reviewed articles sit unpublished because one guard condition in the publishing step had a bug, and nothing alerted anyone. The content was good. The rankings impact was zero, because a ranking system can only reward pages that exist. The lesson cost me weeks: the publishing pipeline needs the same monitoring as the site itself. A success counter that can’t catch a failure is a lie that compounds quietly.

The tradeoff, named

Everything above is slower than buying a “GEO package.” Fixing indexation is real technical work. Extracting unique content requires the owner’s time, every week, indefinitely. There is no version of this where you pay $800/month and your visibility fixes itself while you sleep. Any vendor offering that version is selling the absence of the work, not the work.

What you get for the slower path is an asset that compounds: a site Google can retrieve from, saying things nobody else can say. That’s durable across whatever the results page looks like in 2028.

5 checks to run this afternoon

Search site:yourbusiness.com on Google. Count the results. If pages you care about are missing, or pages you forgot exist are present, that’s your first project.
Load your homepage on your phone, on cellular. Count the seconds. Over 3 is a problem you can feel; Google can feel it too.
Type your domain with and without www. If both load without one redirecting to the other, every page on your site is competing against itself.
Open your most recent blog post and ask: could a competitor paste this on their site unchanged? If yes, it’s commodity content, and it’s doing very little.
Open Search Console → Performance. Impressions from AI Overviews are already counted there. If you’ve never looked, you have no idea whether any of this is working, and neither does your vendor.

One takeaway, since that’s all a piece like this should have: the entry ticket to AI search is the unglamorous technical foundation, and the winning move is publishing what only you know, every week, through a pipeline that actually ships. Google put that in writing. Hold every vendor to it, including us.

Source: Google Search Central, “Optimizing your website for generative AI features on Google Search,” May 2026.