Google's AI search guide: still SEO.

Google ended the AEO debate this week, in a single Search Central doc. The thesis sentence sits a third of the way down, *from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO*. Everything around it is mythbusting, what you do not need to do, what does not exist as a separate discipline, what consultants have been wrong about. The optimization stays what it was. The retrieval is the part that changed.

§01The guide's one-sentence claim

Google published the AI optimization guide inside its Search Central documentation this week, in the Fundamentals section, next to the canonical SEO Starter Guide. The placement is the first signal of the argument. There is no separate generative-AI section. The thing you would do to rank in AI Overviews lives in the same documentation tree as the thing you would do to rank in classical results, because, in Google's framing, it is the same thing.

The argument has a mechanism behind it, not just a marketing line. Google's AI features run on retrieval-augmented generation: the core ranking system fetches the relevant pages, the model generates a response on top of what it retrieved, and the citations Google shows next to the AI output point at the same kinds of pages classical search would have surfaced. Query fan-out, the trick where the system spawns related concurrent queries to broaden the retrieval, runs over the same index. The corpus is the same. The ranking is the same. The thing in the middle that writes a paragraph is new.

§02The mythbusting list

What the guide spends most of its word count on is not-doing. The form is unusual for Google documentation. A list of practices the AI-search-optimization industry has produced over the past two years that Google explicitly says are not required, do not help, and do not exist as a special category.

You do not need a new machine-readable file. The llms.txt proposal that floated through the AI-builder community last year is, in Google's words, not part of how its systems work.

You do not need new markup. There is no special schema.org annotation for AI search. Existing structured data is welcome where it makes sense (recipes, products, articles, events) but is not required for AI-generated answers.

You do not need to chunk your content. Google's systems handle multi-topic pages and read them as humans do. Breaking a page into deliberately AI-sized fragments produces a worse human experience without producing a better AI one.

You do not need to write in a special voice. Content "for AI" is the same content that works for readers, and the model-friendly rewrite that several consulting practices sell is, in Google's framing, optimizing for a target that does not exist.

The aggregate posture, read as one piece, is that the AI search optimization industry as currently practiced is mostly building artifacts for which there is no consumer.

§03RAG is why it is still SEO

The mythbusting only makes sense if you remember what Google's AI features are actually doing under the hood. AI Overviews and the generative answers in Google's results are not a separate model trained on a separate corpus that bypasses the regular Search index. They are a wrapper. A query comes in, the regular ranking system retrieves pages from the regular index, the model conditions on those pages and writes a summary, and the citations point back at the pages that fed the summary. The model contributes synthesis. Retrieval is still where the ranking happens.

That architecture forces the SEO answer. If the pages that show up in AI Overviews are drawn from the same pool of pages classical SEO has always optimized for, then the practices that get a page into the pool, indexability, clean HTML, useful content, are still the practices that decide outcomes. The model on top is not picking favorites it could not see. It is summarizing what retrieval surfaced.

The opposite assumption, which most of the AEO and GEO frameworks implicitly run on, is that the model has some other ingestion path it uses to learn what to cite. It does not. The retrieval is the optimization target. The model is the post-processor.

§04What Google does recommend

The recommendations are short, and most of them are the standard Search Central material: non-commodity content with a perspective, clean indexable HTML, semantic markup where it carries information, reasonable page experience across devices, the usual technical hygiene around duplicate content and canonical URLs. Two recommendations are worth pulling out.

The first is the "unique value, not commodity content" line. The model is good at summarizing facts that exist on a hundred pages. The model is bad at synthesizing perspective that exists on one. The pages that win the citation slot in an AI Overview are the ones that say something the model could not have written from the rest of the index, which is the same thing classical Search has rewarded since the helpful-content updates.

The second is the suggestion to make your site agent-friendly for the emerging agentic experiences. This is Google's quiet acknowledgment that the next traffic-shaped surface to optimize for is not a person at a search box but an agent acting on a person's behalf. The advice is conservative, clear navigation, semantic structure, stable URLs, but the framing matters. The optimization target is shifting from human-clicks-link to agent-fetches-page, and the optimization technique remains, again, foundational SEO.

§05The implication for AI builders

For the audience that ships AI products, the practical move is to stop thinking of GEO and AEO as separate disciplines and start treating "generative AI search" as a new retrieval surface for the same SEO work. The implications cash out in three places.

First, llms.txt becomes optional. The proposal has its uses, a curated entry point for agents crawling your site, a way to expose a structured index where one does not exist, but it should not be confused for AI-search optimization. Google does not consume it for that purpose.

Second, the GEO services market shrinks. Agencies selling "AI-first content rewriting" or "model-targeted schema work" are selling a deliverable Google has now stated it does not consume. Some of the underlying work, writing better content, structuring pages cleanly, fixing crawl issues, is genuinely useful. The framing as a new discipline is not.

Third, the optimization target is the same as it has always been. A useful page, crawlable, indexable, with a perspective that the rest of the index does not duplicate. The retrieval mechanics are new. The thing being retrieved is not.