Google Just Wrote Down How AI Search Works. The Reassuring Bits Are the Warning.

Signpost with arrows pointing in many directions, illustrating the choices businesses face with AI search optimization

A few weeks back I wrote that SEO isn’t dead — it’s entered a new era. Then on May 15th, Google went and published the rulebook for that era.

Quietly. No keynote, no countdown. Just a new “Generative AI fundamentals” section tucked into the Search Central docs, with John Mueller’s name on the announcement.

If you’ve spent the last year getting cold emails about “AEO audits” and “GEO frameworks” and a magic file called llms.txt that was apparently going to save your rankings — this is the document that tells you which of that to keep and which to bin.

I read the whole thing so you don’t have to. Here’s the honest version.

First, the bit that’ll annoy half the consultants in your inbox

There’s been a whole cottage industry built on the idea that AI search is a brand-new game with brand-new rules. New acronyms. New deliverables. New invoices.

Google’s position, in writing now: optimizing for AI search is SEO. AEO, GEO, whatever you want to call it — from Google’s side of the glass, it’s the same search experience and the same fundamentals.

That’s not Google being lazy. AI Overviews and AI Mode are built on the exact ranking systems that already decide who shows up in regular search. So if someone’s selling you “AI optimization” as a separate thing with a separate price tag — fair to ask what’s actually new about it.

How AI search decides what to show (the 30-second version)

Two ideas do the heavy lifting: RAG (the AI builds answers out of real, indexed pages — not thin air) and query fan-out (it quietly runs a handful of related searches and stitches them together).

The practical upshot: your page doesn’t have to match the searcher’s exact words. A genuinely deep page gets pulled in because it answers one of the sub-questions properly. Depth beats exact-match keyword stuffing.

I went deep on both of these — and how ChatGPT and Google AI Mode differ — in the new-era post, so I won’t rehash it here. The headline’s the same: build for the question, not the keyword.

The four things Google says you can stop paying for

This is my favourite part of the guide, because Google rarely names-and-shames specific tactics. When it does, it’s because enough people are selling them that it felt the need to say something out loud.

For Google search specifically, here’s what the guide says you don’t need:

  • llms.txt files. The much-hyped “feed the AI a tidy menu of your site” file. Google may crawl it like any other text file, but it gets zero special treatment. (Other crawlers — Anthropic, OpenAI, Perplexity — may use it, so keep that decision separate. Just don’t do it for Google and expect a payoff.)
  • Chopping content into tiny “AI-friendly” chunks. Google’s systems can read a normal, multi-topic page and pull the relevant bit themselves. Fragmenting your articles just makes them choppier for actual humans.
  • Rewriting copy to be “understood by AI.” It handles synonyms and meaning fine. You don’t need to cram in every long-tail variation like it’s 2012.
  • Special schema or Markdown versions of pages. Keep using structured data for normal rich results — it’s good hygiene — but there’s no secret AI-Overview-unlocking markup. It isn’t a lever.

And one quiet but sharp addition: Google updated its spam policies to confirm they now apply to AI responses too. So the old trick of seeding fake brand mentions across forums to look more authoritative? Same rules, same consequences. Earn the mentions or don’t bother.

Here’s the line everyone skimmed past

Buried in the content section is the whole ballgame, and most recaps glossed right over it. Google draws a line between two kinds of content:

Commodity content — Google’s own example: “7 Tips for First-Time Homebuyers.” Common knowledge. Available from anyone. Nothing in it that didn’t already exist on a hundred other pages.

Non-commodity content — Google’s example: “Why We Waived the Inspection and Saved Money: A Look Inside the Sewer Line.” A specific, lived-through perspective. Real numbers. A decision someone actually made and can explain.

The test underneath it is brutal and simple:

Could ChatGPT produce an equally useful version of this page in 90 seconds?

If yes — that’s commodity content. And commodity content is exactly what AI Overviews are best at swallowing whole and serving up without ever sending anyone to your site.

This is sharper than the old “helpful content” advice, and here’s the difference. “Helpful” asks whether your content is good. “Non-commodity” asks whether it could only have come from you. A first-time-homebuyer guide can be beautifully written, genuinely useful, and rank well — and still be reproducible by a machine in seconds. The quality isn’t the problem. The type is.

Google is telling you, in the politest possible language, that the content most at risk isn’t your bad content. It’s your generic content.

Why this one’s a bit awkward for me to write

Quick confession. Holler runs as an AI-first shop. I’ve got a system of AI agents watching client sites around the clock — uptime, rankings, technical issues, the lot. It’s how one person keeps an eye on what used to need a whole team.

So I spend my days proving AI is brilliant at certain jobs. And here I am telling you the one thing it can’t do is write the content that’ll actually win you AI search visibility.

Both are true. The agents are great at monitoring, pattern-spotting, doing the repetitive grind at 3am. What they can’t do is sit in a client meeting, watch a project nearly go off the rails, and then write up exactly what broke and what we’d do differently. They can’t tell you the real dollar figure on a campaign. They’ve never waived a home inspection.

That’s the moat. Not “AI content bad” — Google’s careful to say it doesn’t penalize AI-assisted work. It’s that volume alone is worthless now. A first-hand account, a number only you have, a real opinion you’ll stand behind — that’s the stuff that can’t be summarized away, because it was never sitting in the training data to begin with.

If you’re a local Edmonton business, don’t skip this

One line in the guide that local and ecommerce folks need to clock: Google explicitly names Google Business Profile and Merchant Center as inputs into AI answers — not just regular search results.

So your GBP isn’t a “set it and forget it” listing anymore. It’s part of the pile Google reaches into when it builds an AI answer about a local service. An outdated profile, a thin product feed, a half-finished listing — that’s a gap quietly costing you visibility in a place you probably aren’t even looking.

If you’ve been meaning to sort your GBP out, this is your nudge. Lower effort than most content work, and it feeds straight into the new system.

What I’d actually do this month

Not a 40-point checklist. Four moves, ordered by bang-for-effort:

  1. Run the 90-second test on your best pages. Pull your top pages by impressions. For each, ask honestly: could ChatGPT write this and lose nothing? The ones that fail are your at-risk pages. Don’t delete them — upgrade them with something only you have: a real client result, your own data, a screenshot, an actual opinion.
  2. Check your high-value pages for nosnippet tags. To show up in AI answers, a page has to be indexed and eligible for a snippet. A stray nosnippet directive — often left over from some old template decision — quietly locks a strong page out entirely. Quick win if you find one.
  3. Consolidate before you create. Got ten thin pages nibbling at slices of the same topic? In this new world, that fragmentation can count against you. One deep, authoritative page usually beats ten shallow ones. Merge, don’t multiply.
  4. Tidy up GBP and any product feeds. See above. Low effort, direct line into AI answers.

And once you’ve done the work — here’s how to actually measure whether it’s paying off, before your boss asks.

The thing Google’s saying without saying it

Most people will read this guide as a pat on the back. See? SEO still works, fundamentals still matter, nothing to panic about. All true.

But that’s not the whole message. The whole message is that the weak, generic, technically-fine-but-soulless content that used to skate by on good formatting? It’s got nowhere to hide now. AI can summarize it and replace it in one move.

The question stopped being “what do we bolt on to optimize for AI.” It’s this:

If your page vanished tomorrow, would anyone actually miss it — or are there ten identical ones already ranking?

That’s the bar now. It’s a higher one. But honestly? It’s the bar good content always should have cleared.


Curious whether your site’s content would survive the 90-second test? That’s basically what I do all day. Book a 15-minute Zoom and I’ll give you a straight answer on where you’re exposed.