For about a year, a whole industry has sprung up selling you "GEO" and "AEO" — generative engine optimisation, answer engine optimisation — courses, audits, and a checklist of new things you simply must do or the AI Overview will forget you exist. This week Google published its own guide to optimising for AI search. It quietly takes that checklist apart, line by line.
I read it the way I read the X algorithm when it went open-source: not for the theory, but for the one question that matters — what do I do differently on Monday morning? Here's how I read it.
"GEO" and "AEO" aren't real — and Google said so in writing
Google's guide addresses the terminology head-on. AEO, GEO, the whole alphabet of new acronyms — Google's position is blunt. From Google Search's perspective, optimising for generative AI search is optimising for the search experience, "and thus still SEO."
There is no separate discipline. There is no second algorithm to game. AI Overviews and AI Mode are built on the same core ranking and quality systems that have ranked pages for two decades. The guide says it plainly: the best practices for SEO "continue to be relevant" because the AI features are rooted in core Search.
That one sentence should save you a lot of money.
The hacks Google just put in the bin
Read the guide as a list of things you can stop doing:
- llms.txt files. Google explicitly says don't create LLMS.txt files, or any "new machine readable files, AI text files, markup, or Markdown". It may crawl them. It does not treat them specially. The llms.txt gold rush is over.
- Chopping content into tiny AI-friendly chunks. No. Google says there's "no ideal page length" and the model understands synonyms and meaning perfectly well. You do not need to pre-digest your pages for a transformer.
- Writing "for AI". Don't adopt a special style for generative search. Don't chase long-tail keyword variations because you've decided the AI needs them spelled out.
- Buying "mentions". Inauthentic mentions sprayed across the web — link-building with a new coat of paint — Google calls out directly as low-value next to genuinely good content.
None of these were ever real optimisations. They were products. Google has just removed the fear they were sold on.
How AI Overviews actually choose content
The mechanism is worth understanding because it kills the mystery. Two parts.
First, retrieval — Google calls it "grounding". The AI uses the core Search ranking systems to retrieve relevant pages from the same index every other result comes from, reads them, generates an answer, and shows clickable links back to the sources.
Second, query fan-out. For a single question, the AI generates a spread of related sub-queries and pulls results for each — ask it how to fix a lawn and it quietly also searches herbicides and weed prevention.
The takeaway is almost boringly simple: if you are indexed and you rank, you are in the pool the AI draws from. If you aren't, no file format saves you. Optimising for AI Overviews is optimising for Search. Same index, same door.
The one thing that genuinely changed
Here's the part that deserves your attention. Google's guide draws a hard line between two kinds of content, and tells you with its own examples which one AI search has no use for.
Commodity content is finished. Google's own example of what not to write is an article called "7 Tips for First-Time Homebuyers". Not penalised, exactly — just pointless. An AI Overview can assemble those seven tips from anywhere, or generate them outright. Restating what is already on the web earns you nothing now.
What it wants instead — again, Google's own example — is a piece called "Why We Waived the Inspection & Saved Money". First-hand experience. A unique point of view. An expert or experienced take that, in Google's words, goes "beyond common knowledge". The thing a model cannot synthesise, because it only exists if someone actually did it and reported back.
That is the real shift, and it isn't a technical one. AI search raises the floor. Generic, competent, derivative content used to earn a respectable ranking. Now it gets absorbed into an AI summary and answered without a click. The only content that reliably survives is content the AI can't write by itself.
Query fan-out means depth beats coverage
One more practical read. Because the AI fans a question out into many sub-questions, shallow coverage gets caught out. One thin page on a subject will match the headline query and miss the ten follow-on queries the fan-out generated. A genuinely thorough piece — one that answers the next questions a real person would actually have — gets pulled into more of those fanned-out results.
Depth was always good SEO. Fan-out just made it pay more.
What I'd actually do Monday morning
- Stop paying for anything with "GEO" or "AEO" in the name. Google has told you in writing it is not a separate discipline. Cancel the course.
- Delete llms.txt from the to-do list. Spend that hour on one good page instead.
- Audit your content for commodity articles. The listicles, the "ultimate guides" that restate the obvious — they're dead weight now. Rewrite the best of them around first-hand experience, and retire the rest.
- Put your own experience into the page. The campaign you actually ran, the number you actually saw, the thing that actually went wrong. That is the part no model can regurgitate — and now it's the part that earns the click.
- Go deep on fewer topics. Answer the follow-on questions, not just the headline one. That's how you catch the fan-out.
- Keep doing the boring technical SEO. Indexable, crawlable, fast, semantic HTML, verified in Search Console. The guide is clear: that is still the price of entry.
The pattern here is the same one we saw when X open-sourced its feed: the platforms are removing the tricks, not adding them. Google didn't publish a new game to play. It published a guide that mostly says stop playing games. Be genuinely useful, from genuine experience, to a genuine person — and you are optimised for AI search by definition.
That's bad news if you were selling GEO courses. It's good news for everyone who was willing to do the real work.
— Tom